def test_grad_of_grad():
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
y = x2 * x2 + x2 * x3
grad_x2, grad_x3 = ad.gradients(y, [x2, x3])
grad_x2_x2, grad_x2_x3 = ad.gradients(grad_x2, [x2, x3])
executor = ad.Executor([y, grad_x2, grad_x3, grad_x2_x2, grad_x2_x3])
x2_val = 2 * np.ones(3)
x3_val = 3 * np.ones(3)
y_val, grad_x2_val, grad_x3_val, grad_x2_x2_val, grad_x2_x3_val = executor.run(
feed_dict={
x2: x2_val,
x3: x3_val
})
expected_yval = x2_val * x2_val + x2_val * x3_val
expected_grad_x2_val = 2 * x2_val + x3_val
expected_grad_x3_val = x2_val
expected_grad_x2_x2_val = 2 * np.ones_like(x2_val)
expected_grad_x2_x3_val = 1 * np.ones_like(x2_val)
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, expected_yval)
assert np.array_equal(grad_x2_val, expected_grad_x2_val)
assert np.array_equal(grad_x3_val, expected_grad_x3_val)
assert np.array_equal(grad_x2_x2_val, expected_grad_x2_x2_val)
assert np.array_equal(grad_x2_x3_val, expected_grad_x2_x3_val)
def test_full_forward_op():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
#ini
ctx = ndarray.gpu(0)
x_val = np.linspace(0, 100, 100).reshape((5, 1, 20))
filters_val = np.ones((1, 1, 20)) * 0.001
y_val = np.zeros((5, 1))
x_val = ndarray.array(x_val, ctx)
filters_val = ndarray.array(filters_val, ctx)
y_val = ndarray.array(y_val, ctx)
outputs = ad.convolution_1d_forward_op(inputs, filters, "NCHW", "VALID", 1)
outputs_pool = ad.pooling_1d_forward_op(outputs, "NCHW", "max", 0, 1, 1)
outputs_relu = ad.activation_forward_op(outputs_pool, "NCHW", "relu")
outputs_f = ad.flatten_op(outputs_relu)
output = ad.fullyactivation_forward_op(outputs_f, "NCHW", "relu")
loss = ad.matmul_op(output, output, trans_A=True) * (1 / 5)
grad_f = ad.gradients(loss, [filters]) #gra返回一个list
executor = ad.Executor([grad_f[0]], ctx=ctx)
g_val = executor.run(feed_dict={
inputs: x_val,
filters: filters_val
}) #返回一个list
print("g_val:", g_val[0].asnumpy())
def test_matmul_two_vars():
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
y = ad.matmul_op(x2, x3)
grad_x2, grad_x3 = ad.gradients(y, [x2, x3])
executor = ad.Executor([y, grad_x2, grad_x3])
x2_val = np.array([[1, 2], [3, 4], [5, 6]]) # 3x2
x3_val = np.array([[7, 8, 9], [10, 11, 12]]) # 2x3
y_val, grad_x2_val, grad_x3_val = executor.run(feed_dict={
x2: x2_val,
x3: x3_val
})
expected_yval = np.matmul(x2_val, x3_val)
expected_grad_x2_val = np.matmul(np.ones_like(expected_yval),
np.transpose(x3_val))
expected_grad_x3_val = np.matmul(np.transpose(x2_val),
np.ones_like(expected_yval))
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, expected_yval)
assert np.array_equal(grad_x2_val, expected_grad_x2_val)
assert np.array_equal(grad_x3_val, expected_grad_x3_val)
def test_add_mul_mix_3():
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
z = x2 * x2 + x2 + x3 + 3
y = z * z + x3
grad_x2, grad_x3 = ad.gradients(y, [x2, x3])
executor = ad.Executor([y, grad_x2, grad_x3])
x2_val = 2 * np.ones(3)
x3_val = 3 * np.ones(3)
y_val, grad_x2_val, grad_x3_val = executor.run(feed_dict={
x2: x2_val,
x3: x3_val
})
z_val = x2_val * x2_val + x2_val + x3_val + 3
expected_yval = z_val * z_val + x3_val
expected_grad_x2_val = 2 * (x2_val * x2_val + x2_val + x3_val +
3) * (2 * x2_val + 1)
expected_grad_x3_val = 2 * (x2_val * x2_val + x2_val + x3_val + 3) + 1
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, expected_yval)
assert np.array_equal(grad_x2_val, expected_grad_x2_val)
assert np.array_equal(grad_x3_val, expected_grad_x3_val)
def test_add_mul_mix_2():
x1 = ad.Variable(name="x1")
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
x4 = ad.Variable(name="x4")
y = x1 + x2 * x3 * x4
grad_x1, grad_x2, grad_x3, grad_x4 = ad.gradients(y, [x1, x2, x3, x4])
executor = ad.Executor([y, grad_x1, grad_x2, grad_x3, grad_x4])
x1_val = 1 * np.ones(3)
x2_val = 2 * np.ones(3)
x3_val = 3 * np.ones(3)
x4_val = 4 * np.ones(3)
y_val, grad_x1_val, grad_x2_val, grad_x3_val, grad_x4_val = executor.run(
feed_dict={
x1: x1_val,
x2: x2_val,
x3: x3_val,
x4: x4_val
})
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, x1_val + x2_val * x3_val * x4_val)
assert np.array_equal(grad_x1_val, np.ones_like(x1_val))
assert np.array_equal(grad_x2_val, x3_val * x4_val)
assert np.array_equal(grad_x3_val, x2_val * x4_val)
assert np.array_equal(grad_x4_val, x2_val * x3_val)
def block_reduction_a(inputs, input_size, blockname=None):
filter2_1_0 = ad.Variable(blockname + "filter2_1_0")
filter2_1_1a = ad.Variable(blockname + "filter2_1_1a")
filter2_1_1b = ad.Variable(blockname + "filter2_1_1b")
filter2_1_1c = ad.Variable(blockname + "filter2_1_1c")
rand = np.random.RandomState(seed=123)
filter2_1_0_val = rand.normal(scale=0.1, size=(384, input_size, 3, 3))
filter2_1_1_vala = rand.normal(scale=0.1, size=(192, input_size, 1, 1))
filter2_1_1_valb = rand.normal(scale=0.1, size=(224, 192, 3, 3))
filter2_1_1_valc = rand.normal(scale=0.1, size=(256, 224, 3, 3))
# branch_0
incep2_1_0 = ad.convolution_2d_forward_op(inputs, filter2_1_0, "NCHW",
"VALID", 2, 2)
# branch 1
incep2_1_1a = ad.convolution_2d_forward_op(inputs, filter2_1_1a, "NCHW",
"SAME", 1, 1)
incep2_1_1b = ad.convolution_2d_forward_op(incep2_1_1a, filter2_1_1b,
"NCHW", "SAME", 1, 1)
incep2_1_1 = ad.convolution_2d_forward_op(incep2_1_1b, filter2_1_1c,
"NCHW", "VALID", 2, 2)
# branch 2
incep2_1_2 = ad.pooling_2d_forward_op(inputs, "NCHW", "mean", 0, 0, 2, 2,
3, 3)
concat2_1a = ad.concat_forward_op(incep2_1_0, incep2_1_1)
concat2_1 = ad.concat_forward_op(concat2_1a, incep2_1_2)
dict = {
filter2_1_0: filter2_1_0_val,
filter2_1_1a: filter2_1_1_vala,
filter2_1_1b: filter2_1_1_valb,
filter2_1_1c: filter2_1_1_valc
}
return concat2_1, dict
def test_lr():
W = ad.Variable(name="W")
b = ad.Variable(name="b")
X = ad.Variable(name="X")
y_ = ad.Variable(name="y_")
ctx = ndarray.gpu(0)
# ini
x_val = np.linspace(0, 1, 100).reshape((100, 1))
y_val = x_val + 0.5
W_val = np.array([[0.1]])
b_val = np.array([0.1])
x_val = ndarray.array(x_val, ctx)
W_val = ndarray.array(W_val, ctx)
b_val = ndarray.array(b_val, ctx)
y_val = ndarray.array(y_val, ctx)
z = ad.matmul_op(X, W)
# z.shape = (100,1)
# b.shape = (1,1)
y = z + ad.broadcastto_op(b, z)
# y = (100,1)
y = ad.fullyactivation_forward_op(y, "NCHW", "relu")
loss = ad.matmul_op(y + (-1) * y_, y + (-1) * y_, trans_A=True) * (1 / 100)
# loss = ad.softmaxcrossentropy_op(y, y_)
grad_W, grad_b = ad.gradients(loss, [W, b])
executor = ad.Executor([loss, grad_W, grad_b], ctx)
aph = 1e-6
for i in range(100):
loss_val, grad_W_val, grad_b_val = executor.run(feed_dict={
X: x_val,
b: b_val,
W: W_val,
y_: y_val
})
grad_W_val = grad_W_val.asnumpy()
W_val = W_val.asnumpy()
W_val = W_val - aph * grad_W_val
W_val = ndarray.array(W_val, ctx)
grad_b_val = grad_b_val.asnumpy()
b_val = b_val.asnumpy()
b_val = b_val - aph * grad_b_val
b_val = ndarray.array(b_val, ctx)
print(W_val.asnumpy(), b_val.asnumpy())
def test_l1_l2_regular():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
# ini
ctx = ndarray.gpu(0)
x_val = np.ones((5, 2)) * 0.5
x_val = ndarray.array(x_val, ctx)
loss = ad.l2regular_op(inputs)
# loss = ad.l1regular_op(inputs)
grad_f = ad.gradients(loss, [inputs]) # gra返回一个list
executor = ad.Executor([loss, grad_f[0]], ctx=ctx)
g_val = executor.run(feed_dict={inputs: x_val}) # 返回一个list
print("g_val:", g_val[0].asnumpy())
print("g_val:", g_val[1].asnumpy())
def convolutional_block(inputs,
kernel_size,
in_filter,
out_filters,
block_name,
stride=1):
f1, f2, f3 = out_filters
W1 = ad.Variable(block_name + "W1")
W2 = ad.Variable(block_name + "W2")
W3 = ad.Variable(block_name + "W3")
W_shortcut = ad.Variable(block_name + "W_shortcut")
rand = np.random.RandomState(seed=123)
W1_val = rand.normal(scale=0.1, size=(f1, in_filter, 1, 1))
W2_val = rand.normal(scale=0.1, size=(f2, f1, kernel_size, kernel_size))
W3_val = rand.normal(scale=0.1, size=(f3, f2, 1, 1))
W_shortcut_val = rand.normal(scale=0.1, size=(in_filter, f3, 1, 1))
# conv1
conv1 = ad.convolution_2d_forward_op(inputs, W1, "NCHW", "SAME", stride,
stride)
bn1 = ad.bn_forward_op(conv1, "NCHW", "pre_activation")
act1 = ad.activation_forward_op(bn1, "NCHW", "relu")
# conv2
conv2 = ad.convolution_2d_forward_op(act1, W2, "NCHW", "SAME", stride,
stride)
bn2 = ad.bn_forward_op(conv2, "NCHW", "pre_activation")
act2 = ad.activation_forward_op(bn2, "NCHW", "relu")
# conv3
conv3 = ad.convolution_2d_forward_op(act2, W3, "NCHW", "VALID", stride,
stride)
bn3 = ad.bn_forward_op(conv3, "NCHW", "pre_activation")
#shortcut_path
conv4 = ad.convolution_2d_forward_op(inputs, W_shortcut, "NCHW", "VALID",
stride, stride)
shortcut = ad.bn_forward_op(conv4, "NCHW", "pre_activation")
# shortcut
add = ad.add_op(bn3, shortcut)
act4 = ad.activation_forward_op(add, "NCHW", "relu")
dict = {W1: W1_val, W2: W2_val, W3: W3_val, W_shortcut: W_shortcut_val}
return act4, dict
def test_mul_two_vars():
x2 = ad.Variable(name="x2")
x3 = ad.Variable(name="x3")
y = x2 * x3
grad_x2, grad_x3 = ad.gradients(y, [x2, x3])
executor = ad.Executor([y, grad_x2, grad_x3])
x2_val = 2 * np.ones(3)
x3_val = 3 * np.ones(3)
y_val, grad_x2_val, grad_x3_val = executor.run(feed_dict={
x2: x2_val,
x3: x3_val
})
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, x2_val * x3_val)
assert np.array_equal(grad_x2_val, x3_val)
assert np.array_equal(grad_x3_val, x2_val)
def test_exp_grad():
x = ad.Variable("x")
y = ad.exp_op(x)
x_grad, = ad.gradients(y, [x])
executor = ad.Executor([y, x_grad])
x_val = 1
y_val, x_grad_val = executor.run(feed_dict={x: x_val})
print(y_val)
print(x_grad_val)
def test_l1_l2_cross_loss():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
# ini
ctx = ndarray.gpu(0)
x_val = np.ones((5, 2)) * 0.5
filters_val = np.ones((2, 2, 10)) * 0.001
y_val = np.ones((5, 2))
x_val = ndarray.array(x_val, ctx)
filters_val = ndarray.array(filters_val, ctx)
y_val = ndarray.array(y_val, ctx)
# loss = ad.crossEntropy_op(inputs, y_)
loss = ad.l1loss_op(inputs, y_)
grad_f = ad.gradients(loss, [inputs, y_]) # gra返回一个list
executor = ad.Executor([loss, grad_f[0], grad_f[1]], ctx=ctx)
g_val = executor.run(feed_dict={inputs: x_val, y_: y_val}) # 返回一个list
print("g_val:", g_val[0].asnumpy())
print("g_val:", g_val[1].asnumpy())
print("g_val:", g_val[2].asnumpy())
def test_convolution_3d_forward_op():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
#ini
ctx = ndarray.gpu(0)
x_val = np.linspace(0, 100, 135).reshape((5, 1, 3, 3, 3))
filters_val = np.ones((1, 1, 2, 2, 2)) * 0.001
y_val = np.zeros((5, 1))
x_val = ndarray.array(x_val, ctx)
filters_val = ndarray.array(filters_val, ctx)
y_val = ndarray.array(y_val, ctx)
outputs = ad.convolution_3d_forward_op(inputs, filters, "NCHW", "VALID", 1,
1, 1)
outputs_pool = ad.pooling_3d_forward_op(outputs, "NCHW", "max", 0, 0, 0, 1,
1, 1, 2, 2, 2)
outputs_relu = ad.activation_forward_op(outputs_pool, "NCHW", "relu")
outputs_dro = ad.dropout_forward_op(outputs_relu, "NCHW", 0.5, 0)
outputs_f = ad.flatten_op(outputs_dro)
loss = ad.matmul_op(outputs_f, outputs_f, trans_A=True) * (1 / 5)
grad_inputs, grad_f = ad.gradients(loss, [inputs, filters])
executor = ad.Executor([loss, grad_f], ctx=ctx)
aph = 1.0e-6
for i in range(20):
loss_val, filters_grad_val = executor.run(feed_dict={
inputs: x_val,
filters: filters_val
})
filters_val = filters_val.asnumpy()
filters_grad_val = filters_grad_val.asnumpy()
filters_val = filters_val - aph * filters_grad_val
filters_val = ndarray.array(filters_val, ctx)
print("loss_val:", loss_val.asnumpy())
print("filters_val:", filters_val.asnumpy())
def test_exp():
x1 = ad.Variable("x1")
x2 = ad.exp_op(x1)
x3 = x2 + 1
x4 = x2 * x3
x1_grad, = ad.gradients(x4, [x1])
executor = ad.Executor([x4])
x1_val = 1
x4_val, x1_grad = executor.run(feed_dict={x1: x1_val})
print(x4_val)
print(x1_grad)
def test_identity():
x2 = ad.Variable(name="x2")
y = x2
grad_x2, = ad.gradients(y, [x2])
executor = ad.Executor([y, grad_x2])
x2_val = 2 * np.ones(3)
y_val, grad_x2_val = executor.run(feed_dict={x2: x2_val})
assert isinstance(y, ad.Node)
assert np.array_equal(y_val, x2_val)
assert np.array_equal(grad_x2_val, np.ones_like(x2_val))
def block_inception_a(inputs, blockname=None):
filter1_1_0 = ad.Variable(blockname + "filter1_1_0")
filter1_1_1a = ad.Variable(blockname + "filter1_1_1a")
filter1_1_1b = ad.Variable(blockname + "filter1_1_1b")
filter1_1_2a = ad.Variable(blockname + "filter1_1_2a")
filter1_1_2b = ad.Variable(blockname + "filter1_1_2b")
filter1_1_2c = ad.Variable(blockname + "filter1_1_2c")
filter1_1_3 = ad.Variable(blockname + "filter1_1_3a")
rand = np.random.RandomState(seed=123)
filter1_1_0_val = rand.normal(scale=0.1, size=(96, 384, 1, 1))
filter1_1_1_vala = rand.normal(scale=0.1, size=(64, 384, 1, 1))
filter1_1_1_valb = rand.normal(scale=0.1, size=(96, 64, 3, 3))
filter1_1_2_vala = rand.normal(scale=0.1, size=(64, 384, 1, 1))
filter1_1_2_valb = rand.normal(scale=0.1, size=(96, 64, 3, 3))
filter1_1_2_valc = rand.normal(scale=0.1, size=(96, 96, 3, 3))
filter1_1_3_val = rand.normal(scale=0.1, size=(96, 384, 1, 1))
# branch_0
incep1_1_0 = ad.convolution_2d_forward_op(inputs, filter1_1_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep1_1_1a = ad.convolution_2d_forward_op(inputs, filter1_1_1a, "NCHW",
"SAME", 1, 1)
incep1_1_1 = ad.convolution_2d_forward_op(incep1_1_1a, filter1_1_1b,
"NCHW", "SAME", 1, 1)
# branch 2
incep1_1_2a = ad.convolution_2d_forward_op(inputs, filter1_1_2a, "NCHW",
"SAME", 1, 1)
incep1_1_2b = ad.convolution_2d_forward_op(incep1_1_2a, filter1_1_2b,
"NCHW", "SAME", 1, 1)
incep1_1_2 = ad.convolution_2d_forward_op(incep1_1_2b, filter1_1_2c,
"NCHW", "SAME", 1, 1)
# branch 3
incep1_1_3a = ad.pooling_2d_forward_op(inputs, "NCHW", "mean", 1, 1, 1, 1,
3, 3)
incep1_1_3 = ad.convolution_2d_forward_op(incep1_1_3a, filter1_1_3, "NCHW",
"SAME", 1, 1)
concat1_1a = ad.concat_forward_op(incep1_1_0, incep1_1_1)
concat1_1b = ad.concat_forward_op(concat1_1a, incep1_1_2)
concat1_1 = ad.concat_forward_op(concat1_1b, incep1_1_3)
dict = {
filter1_1_0: filter1_1_0_val,
filter1_1_1a: filter1_1_1_vala,
filter1_1_1b: filter1_1_1_valb,
filter1_1_2a: filter1_1_2_vala,
filter1_1_2b: filter1_1_2_valb,
filter1_1_2c: filter1_1_2_valc,
filter1_1_3: filter1_1_3_val
}
return concat1_1, dict
def test_sigmoid_conv_1d():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
# ini
ctx = ndarray.gpu(0)
x_val = np.linspace(0, 100, 80).reshape((5, 1, 4, 4))
filters_val = np.ones((1, 1, 3, 3)) * 0.001
y_val = np.zeros((5, 1))
x_val = ndarray.array(x_val, ctx)
filters_val = ndarray.array(filters_val, ctx)
y_val = ndarray.array(y_val, ctx)
outputs = ad.convolution_2d_forward_op(inputs, filters, "NCHW", "VALID", 1,
1)
# outputs_pool = ad.pooling_2d_forward_op(outputs, "NCHW", "max", 0, 0, 1, 1, 2, 2)
outputs_relu = ad.activation_forward_op(outputs, "NCHW", "relu")
executor = ad.Executor([outputs_relu], ctx=ctx)
loss_val = executor.run(feed_dict={inputs: x_val, filters: filters_val})
print("loss_val:", loss_val[0].asnumpy())
def test_exp_log_reverse_pow():
inputs = ad.Variable("inputs")
filters = ad.Variable("filters")
y_ = ad.Variable(name="y_")
# ini
ctx = ndarray.gpu(0)
x_val = np.linspace(0, 100, 80).reshape((5, 1, 4, 4))
filters_val = np.ones((1, 1, 3, 3)) * 0.001
y_val = np.zeros((5, 1))
x_val = ndarray.array(x_val, ctx)
filters_val = ndarray.array(filters_val, ctx)
y_val = ndarray.array(y_val, ctx)
#outputs = ad.exp_op(inputs)
#outputs = ad.log_op(inputs)
#outputs = ad.reverse_op(inputs)
outputs = ad.pow_op(inputs, 2)
grad_out = ad.gradients(outputs, [inputs])
executor = ad.Executor([outputs, grad_out[0]], ctx=ctx)
result = executor.run(feed_dict={inputs: filters_val})
print(result[0].asnumpy())
print(result[1].asnumpy())
def test_reduce_mean():
inputs = ad.Variable("inputs")
ctx = ndarray.gpu(0)
shape = (2, 2, 3)
x = np.random.uniform(0, 20, shape).astype(np.float32)
arr_x = ndarray.array(x, ctx=ctx)
outputs = ad.reduce_mean_op(inputs, 1)
f_out = ad.pow_op(outputs, 2)
grad_out = ad.gradients(f_out, [inputs])
executor = ad.Executor([outputs, f_out, grad_out[0]], ctx=ctx)
result = executor.run(feed_dict={inputs: arr_x})
print(arr_x.asnumpy())
print(result[0].asnumpy())
print(result[1].asnumpy())
print(result[2].asnumpy())
def block_reduction_b(inputs, input_size, blockname=None):
filter3_1_0a = ad.Variable(blockname + "filter3_1_0a")
filter3_1_0b = ad.Variable(blockname + "filter3_1_0b")
filter3_1_1a = ad.Variable(blockname + "filter3_1_1a")
filter3_1_1b = ad.Variable(blockname + "filter3_1_1b")
filter3_1_1c = ad.Variable(blockname + "filter3_1_1c")
filter3_1_1d = ad.Variable(blockname + "filter3_1_1d")
rand = np.random.RandomState(seed=123)
filter3_1_0_vala = rand.normal(scale=0.1, size=(192, input_size, 1, 1))
filter3_1_0_valb = rand.normal(scale=0.1, size=(192, 192, 3, 3))
filter3_1_1_vala = rand.normal(scale=0.1, size=(256, input_size, 1, 1))
filter3_1_1_valb = rand.normal(scale=0.1, size=(256, 256, 1, 7))
filter3_1_1_valc = rand.normal(scale=0.1, size=(320, 256, 7, 1))
filter3_1_1_vald = rand.normal(scale=0.1, size=(320, 320, 3, 3))
# branch_0
incep3_1_0a = ad.convolution_2d_forward_op(inputs, filter3_1_0a, "NCHW",
"SAME", 1, 1)
incep3_1_0 = ad.convolution_2d_forward_op(incep3_1_0a, filter3_1_0b,
"NCHW", "VALID", 2, 2)
# branch 1
incep3_1_1a = ad.convolution_2d_forward_op(inputs, filter3_1_1a, "NCHW",
"SAME", 1, 1)
incep3_1_1b = ad.convolution_2d_forward_op(incep3_1_1a, filter3_1_1b,
"NCHW", "SAME", 1, 1)
incep3_1_1c = ad.convolution_2d_forward_op(incep3_1_1b, filter3_1_1c,
"NCHW", "SAME", 1, 1)
incep3_1_1 = ad.convolution_2d_forward_op(incep3_1_1c, filter3_1_1d,
"NCHW", "VALID", 2, 2)
# branch 2
incep3_1_2 = ad.pooling_2d_forward_op(inputs, "NCHW", "mean", 0, 0, 2, 2,
3, 3)
concat3_1a = ad.concat_forward_op(incep3_1_0, incep3_1_1)
concat3_1 = ad.concat_forward_op(concat3_1a, incep3_1_2)
dict = {
filter3_1_0a: filter3_1_0_vala,
filter3_1_0b: filter3_1_0_valb,
filter3_1_1a: filter3_1_1_vala,
filter3_1_1b: filter3_1_1_valb,
filter3_1_1c: filter3_1_1_valc,
filter3_1_1d: filter3_1_1_vald
}
return concat3_1, dict
def inception_v3():
X = ad.Placeholder("inputs")
y_ = ad.Placeholder("y_")
inputs = ad.Variable("inputs")
filterb_1 = ad.Variable("filterb_1")
filterb_2 = ad.Variable("filterb_2")
filterb_3 = ad.Variable("filterb_3")
filterb_4 = ad.Variable("filterb_4")
filterb_5 = ad.Variable("filterb_5")
ctx = ndarray.gpu(0)
filtersb_val1 = np.ones((32, 3, 3, 3)) * 0.01
filtersb_val2 = np.ones((32, 32, 3, 3)) * 0.01
filtersb_val3 = np.ones((64, 32, 3, 3)) * 0.01
filtersb_val4 = np.ones((80, 64, 1, 1)) * 0.01
filtersb_val5 = np.ones((192, 80, 3, 3)) * 0.01
y_val = np.zeros((147, 147, 64))
#inception前
covb_1 = ad.convolution_2d_forward_op(inputs, filterb_1, "NCHW", "VALID",
2, 2)
covb_2 = ad.convolution_2d_forward_op(covb_1, filterb_2, "NCHW", "VALID",
1, 1)
covb_3 = ad.convolution_2d_forward_op(covb_2, filterb_3, "NCHW", "SAME", 1,
1)
poolb = ad.pooling_2d_forward_op(covb_3, "NCHW", "max", 0, 0, 2, 2, 3, 3)
covb_4 = ad.convolution_2d_forward_op(poolb, filterb_4, "NCHW", "VALID", 1,
1)
covb_5 = ad.convolution_2d_forward_op(covb_4, filterb_5, "NCHW", "VALID",
1, 1)
covb = ad.pooling_2d_forward_op(covb_5, "NCHW", "max", 0, 0, 2, 2, 3, 3)
# inception_moudle1
#inception_moudle1_1
filter1_1_0 = ad.Variable("filter1_1_0")
filter1_1_1a = ad.Variable("filter1_1_1a")
filter1_1_1b = ad.Variable("filter1_1_1b")
filter1_1_2a = ad.Variable("filter1_1_2a")
filter1_1_2b = ad.Variable("filter1_1_2b")
filter1_1_2c = ad.Variable("filter1_1_2c")
filter1_1_3 = ad.Variable("filter1_1_3a")
filter1_1_0_val = np.ones((64, 192, 1, 1)) * 0.01
filter1_1_1_vala = np.ones((48, 192, 1, 1)) * 0.01
filter1_1_1_valb = np.ones((64, 48, 5, 5)) * 0.01
filter1_1_2_vala = np.ones((64, 192, 1, 1)) * 0.01
filter1_1_2_valb = np.ones((96, 64, 3, 3)) * 0.01
filter1_1_2_valc = np.ones((96, 96, 3, 3)) * 0.01
filter1_1_3_val = np.ones((32, 192, 1, 1)) * 0.01
# branch_0
incep1_1_0 = ad.convolution_2d_forward_op(covb, filter1_1_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep1_1_1a = ad.convolution_2d_forward_op(covb, filter1_1_1a, "NCHW",
"SAME", 1, 1)
incep1_1_1 = ad.convolution_2d_forward_op(incep1_1_1a, filter1_1_1b,
"NCHW", "SAME", 1, 1)
# branch 2
incep1_1_2a = ad.convolution_2d_forward_op(covb, filter1_1_2a, "NCHW",
"SAME", 1, 1)
incep1_1_2b = ad.convolution_2d_forward_op(incep1_1_2a, filter1_1_2b,
"NCHW", "SAME", 1, 1)
incep1_1_2 = ad.convolution_2d_forward_op(incep1_1_2b, filter1_1_2c,
"NCHW", "SAME", 1, 1)
# branch 3
incep1_1_3a = ad.pooling_2d_forward_op(covb, "NCHW", "mean", 1, 1, 1, 1, 3,
3)
incep1_1_3 = ad.convolution_2d_forward_op(incep1_1_3a, filter1_1_3, "NCHW",
"SAME", 1, 1)
concat1_1a = ad.concat_forward_op(incep1_1_0, incep1_1_1)
concat1_1b = ad.concat_forward_op(concat1_1a, incep1_1_2)
concat1_1 = ad.concat_forward_op(concat1_1b, incep1_1_3)
#inception_moudle1_2
filter1_2_0 = ad.Variable("filter1_2_0")
filter1_2_1a = ad.Variable("filter1_2_1a")
filter1_2_1b = ad.Variable("filter1_2_1b")
filter1_2_2a = ad.Variable("filter1_2_2a")
filter1_2_2b = ad.Variable("filter1_2_2b")
filter1_2_2c = ad.Variable("filter1_2_2c")
filter1_2_3 = ad.Variable("filter1_2_3a")
filter1_2_0_val = np.ones((64, 256, 1, 1)) * 0.01
filter1_2_1_vala = np.ones((48, 256, 1, 1)) * 0.01
filter1_2_1_valb = np.ones((64, 48, 5, 5)) * 0.01
filter1_2_2_vala = np.ones((64, 256, 1, 1)) * 0.01
filter1_2_2_valb = np.ones((96, 64, 3, 3)) * 0.01
filter1_2_2_valc = np.ones((96, 96, 3, 3)) * 0.01
filter1_2_3_val = np.ones((64, 256, 1, 1)) * 0.01
# branch_0
incep1_2_0 = ad.convolution_2d_forward_op(concat1_1, filter1_2_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep1_2_1a = ad.convolution_2d_forward_op(concat1_1, filter1_2_1a, "NCHW",
"SAME", 1, 1)
incep1_2_1 = ad.convolution_2d_forward_op(incep1_2_1a, filter1_2_1b,
"NCHW", "SAME", 1, 1)
# branch 2
incep1_2_2a = ad.convolution_2d_forward_op(concat1_1, filter1_2_2a, "NCHW",
"SAME", 1, 1)
incep1_2_2b = ad.convolution_2d_forward_op(incep1_2_2a, filter1_2_2b,
"NCHW", "SAME", 1, 1)
incep1_2_2 = ad.convolution_2d_forward_op(incep1_2_2b, filter1_2_2c,
"NCHW", "SAME", 1, 1)
# branch 3
incep1_2_3a = ad.pooling_2d_forward_op(concat1_1, "NCHW", "mean", 1, 1, 1,
1, 3, 3)
incep1_2_3 = ad.convolution_2d_forward_op(incep1_2_3a, filter1_2_3, "NCHW",
"SAME", 1, 1)
concat1_2a = ad.concat_forward_op(incep1_2_0, incep1_2_1)
concat1_2b = ad.concat_forward_op(concat1_2a, incep1_2_2)
concat1_2 = ad.concat_forward_op(concat1_2b, incep1_2_3)
# inception_moudle1_3
filter1_3_0 = ad.Variable("filter1_3_0")
filter1_3_1a = ad.Variable("filter1_3_1a")
filter1_3_1b = ad.Variable("filter1_3_1b")
filter1_3_2a = ad.Variable("filter1_3_2a")
filter1_3_2b = ad.Variable("filter1_3_2b")
filter1_3_2c = ad.Variable("filter1_3_2c")
filter1_3_3 = ad.Variable("filter1_3_3a")
filter1_3_0_val = np.ones((64, 288, 1, 1)) * 0.01
filter1_3_1_vala = np.ones((48, 288, 1, 1)) * 0.01
filter1_3_1_valb = np.ones((64, 48, 5, 5)) * 0.01
filter1_3_2_vala = np.ones((64, 288, 1, 1)) * 0.01
filter1_3_2_valb = np.ones((96, 64, 3, 3)) * 0.01
filter1_3_2_valc = np.ones((96, 96, 3, 3)) * 0.01
filter1_3_3_val = np.ones((64, 288, 1, 1)) * 0.01
# branch_0
incep1_3_0 = ad.convolution_2d_forward_op(concat1_2, filter1_3_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep1_3_1a = ad.convolution_2d_forward_op(concat1_2, filter1_3_1a, "NCHW",
"SAME", 1, 1)
incep1_3_1 = ad.convolution_2d_forward_op(incep1_3_1a, filter1_3_1b,
"NCHW", "SAME", 1, 1)
# branch 2
incep1_3_2a = ad.convolution_2d_forward_op(concat1_2, filter1_3_2a, "NCHW",
"SAME", 1, 1)
incep1_3_2b = ad.convolution_2d_forward_op(incep1_3_2a, filter1_3_2b,
"NCHW", "SAME", 1, 1)
incep1_3_2 = ad.convolution_2d_forward_op(incep1_3_2b, filter1_3_2c,
"NCHW", "SAME", 1, 1)
# branch 3
incep1_3_3a = ad.pooling_2d_forward_op(concat1_2, "NCHW", "mean", 1, 1, 1,
1, 3, 3)
incep1_3_3 = ad.convolution_2d_forward_op(incep1_3_3a, filter1_3_3, "NCHW",
"SAME", 1, 1)
concat1_3a = ad.concat_forward_op(incep1_3_0, incep1_3_1)
concat1_3b = ad.concat_forward_op(concat1_3a, incep1_3_2)
concat1_3 = ad.concat_forward_op(concat1_3b, incep1_3_3)
#
#
#
#
# # inception_moudle2
# inception_moudle2_1
filter2_1_0 = ad.Variable("filter2_1_0")
filter2_1_1a = ad.Variable("filter2_1_1a")
filter2_1_1b = ad.Variable("filter2_1_1b")
filter2_1_1c = ad.Variable("filter2_1_1c")
filter2_1_0_val = np.ones((384, 288, 3, 3)) * 0.01
filter2_1_1_vala = np.ones((64, 288, 1, 1)) * 0.01
filter2_1_1_valb = np.ones((96, 64, 3, 3)) * 0.01
filter2_1_1_valc = np.ones((96, 96, 3, 3)) * 0.01
# branch_0
incep2_1_0 = ad.convolution_2d_forward_op(concat1_3, filter2_1_0, "NCHW",
"VALID", 2, 2)
# branch 1
incep2_1_1a = ad.convolution_2d_forward_op(concat1_3, filter2_1_1a, "NCHW",
"SAME", 1, 1)
incep2_1_1b = ad.convolution_2d_forward_op(incep2_1_1a, filter2_1_1b,
"NCHW", "SAME", 1, 1)
incep2_1_1 = ad.convolution_2d_forward_op(incep2_1_1b, filter2_1_1c,
"NCHW", "VALID", 2, 2)
# branch 2
incep2_1_2 = ad.pooling_2d_forward_op(concat1_3, "NCHW", "mean", 0, 0, 2,
2, 3, 3)
concat2_1a = ad.concat_forward_op(incep2_1_0, incep2_1_1)
concat2_1 = ad.concat_forward_op(concat2_1a, incep2_1_2)
# inception_moudle2_2
filter2_2_0 = ad.Variable("filter2_2_0")
filter2_2_1a = ad.Variable("filter2_2_1a")
filter2_2_1b = ad.Variable("filter2_2_1b")
filter2_2_1c = ad.Variable("filter2_2_1c")
filter2_2_2a = ad.Variable("filter2_2_2a")
filter2_2_2b = ad.Variable("filter2_2_2b")
filter2_2_2c = ad.Variable("filter2_2_2c")
filter2_2_2d = ad.Variable("filter2_2_2d")
filter2_2_2e = ad.Variable("filter2_2_2e")
filter2_2_3 = ad.Variable("filter2_2_3a")
filter2_2_0_val = np.ones((192, 768, 1, 1)) * 0.01
filter2_2_1_vala = np.ones((128, 768, 1, 1)) * 0.01
filter2_2_1_valb = np.ones((128, 128, 1, 7)) * 0.01
filter2_2_1_valc = np.ones((192, 128, 7, 1)) * 0.01
filter2_2_2_vala = np.ones((128, 768, 1, 1)) * 0.01
filter2_2_2_valb = np.ones((128, 128, 7, 1)) * 0.01
filter2_2_2_valc = np.ones((128, 128, 1, 7)) * 0.01
filter2_2_2_vald = np.ones((128, 128, 7, 1)) * 0.01
filter2_2_2_vale = np.ones((192, 128, 1, 7)) * 0.01
filter2_2_3_val = np.ones((192, 768, 1, 1)) * 0.01
# branch_0
incep2_2_0 = ad.convolution_2d_forward_op(concat2_1, filter2_2_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep2_2_1a = ad.convolution_2d_forward_op(concat2_1, filter2_2_1a, "NCHW",
"SAME", 1, 1)
incep2_2_1b = ad.convolution_2d_forward_op(incep2_2_1a, filter2_2_1b,
"NCHW", "SAME", 1, 1)
incep2_2_1 = ad.convolution_2d_forward_op(incep2_2_1b, filter2_2_1c,
"NCHW", "SAME", 1, 1)
# branch 2
incep2_2_2a = ad.convolution_2d_forward_op(concat2_1, filter2_2_2a, "NCHW",
"SAME", 1, 1)
incep2_2_2b = ad.convolution_2d_forward_op(incep2_2_2a, filter2_2_2b,
"NCHW", "SAME", 1, 1)
incep2_2_2c = ad.convolution_2d_forward_op(incep2_2_2b, filter2_2_2c,
"NCHW", "SAME", 1, 1)
incep2_2_2d = ad.convolution_2d_forward_op(incep2_2_2c, filter2_2_2d,
"NCHW", "SAME", 1, 1)
incep2_2_2 = ad.convolution_2d_forward_op(incep2_2_2d, filter2_2_2e,
"NCHW", "SAME", 1, 1)
# branch 3
incep2_2_3a = ad.pooling_2d_forward_op(concat2_1, "NCHW", "mean", 1, 1, 1,
1, 3, 3)
incep2_2_3 = ad.convolution_2d_forward_op(incep2_2_3a, filter2_2_3, "NCHW",
"SAME", 1, 1)
concat2_2a = ad.concat_forward_op(incep2_2_0, incep2_2_1)
concat2_2b = ad.concat_forward_op(concat2_2a, incep2_2_2)
concat2_2 = ad.concat_forward_op(concat2_2b, incep2_2_3)
# inception_moudle2_3
# filter2_3_0 = ad.Variable("filter2_3_0")
# filter2_3_1a = ad.Variable("filter2_3_1a")
# filter2_3_1b = ad.Variable("filter2_3_1b")
# filter2_3_1c = ad.Variable("filter2_3_1c")
# filter2_3_2a = ad.Variable("filter2_3_2a")
# filter2_3_2b = ad.Variable("filter2_3_2b")
# filter2_3_2c = ad.Variable("filter2_3_2c")
# filter2_3_2d = ad.Variable("filter2_3_2d")
# filter2_3_2e = ad.Variable("filter2_3_2e")
# filter2_3_3 = ad.Variable("filter2_3_3a")
#
# filter2_3_0_val = np.ones((192, 768, 1, 1)) * 0.01
# filter2_3_1_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_3_1_valb = np.ones((160, 160, 1, 7)) * 0.01
# filter2_3_1_valc = np.ones((192, 160, 7, 1)) * 0.01
# filter2_3_2_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_3_2_valb = np.ones((160, 160, 7,1)) * 0.01
# filter2_3_2_valc = np.ones((160, 160, 1, 7)) * 0.01
# filter2_3_2_vald = np.ones((160, 160, 7, 1)) * 0.01
# filter2_3_2_vale = np.ones((192, 160, 1, 7)) * 0.01
# filter2_3_3_val = np.ones((192, 768, 1, 1)) * 0.01
#
# # branch_0
# incep2_3_0 = ad.convolution_2d_forward_op(concat2_2, filter2_3_0, "NCHW", "SAME", 1, 1)
# # branch 1
# incep2_3_1a = ad.convolution_2d_forward_op(concat2_2, filter2_3_1a, "NCHW", "SAME", 1, 1)
# incep2_3_1b = ad.convolution_2d_forward_op(incep2_3_1a, filter2_3_1b, "NCHW", "SAME", 1, 1)
# incep2_3_1 = ad.convolution_2d_forward_op(incep2_3_1b, filter2_3_1c, "NCHW", "SAME", 1, 1)
# # branch 2
# incep2_3_2a = ad.convolution_2d_forward_op(concat2_2, filter2_3_2a, "NCHW", "SAME", 1, 1)
# incep2_3_2b = ad.convolution_2d_forward_op(incep2_3_2a, filter2_3_2b, "NCHW", "SAME", 1, 1)
# incep2_3_2c = ad.convolution_2d_forward_op(incep2_3_2b, filter2_3_2c, "NCHW", "SAME", 1, 1)
# incep2_3_2d = ad.convolution_2d_forward_op(incep2_3_2c, filter2_3_2d, "NCHW", "SAME", 1, 1)
# incep2_3_2 = ad.convolution_2d_forward_op(incep2_3_2d, filter2_3_2e, "NCHW", "SAME", 1, 1)
# # branch 3
# incep2_3_3a = ad.pooling_2d_forward_op(concat2_2, "NCHW", "mean", 1, 1, 1, 1, 3, 3)
# incep2_3_3 = ad.convolution_2d_forward_op(incep2_3_3a, filter2_3_3, "NCHW", "SAME", 1, 1)
#
# concat2_3a = ad.concat_forward_op(incep2_3_0, incep2_3_1)
# concat2_3b = ad.concat_forward_op(concat2_3a, incep2_3_2)
# concat2_3 = ad.concat_forward_op(concat2_3b, incep2_3_3)
# inception_moudle2_4
# filter2_4_0 = ad.Variable("filter2_4_0")
# filter2_4_1a = ad.Variable("filter2_4_1a")
# filter2_4_1b = ad.Variable("filter2_4_1b")
# filter2_4_1c = ad.Variable("filter2_4_1c")
# filter2_4_2a = ad.Variable("filter2_4_2a")
# filter2_4_2b = ad.Variable("filter2_4_2b")
# filter2_4_2c = ad.Variable("filter2_4_2c")
# filter2_4_2d = ad.Variable("filter2_4_2d")
# filter2_4_2e = ad.Variable("filter2_4_2e")
# filter2_4_3 = ad.Variable("filter2_4_3a")
#
# filter2_4_0_val = np.ones((192, 768, 1, 1)) * 0.01
# filter2_4_1_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_4_1_valb = np.ones((160, 160, 1, 7)) * 0.01
# filter2_4_1_valc = np.ones((192, 160, 7, 1)) * 0.01
# filter2_4_2_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_4_2_valb = np.ones((160, 160, 7, 1)) * 0.01
# filter2_4_2_valc = np.ones((160, 160, 1, 7)) * 0.01
# filter2_4_2_vald = np.ones((160, 160, 7, 1)) * 0.01
# filter2_4_2_vale = np.ones((192, 160, 1, 7)) * 0.01
# filter2_4_3_val = np.ones((192, 768, 1, 1)) * 0.01
#
# # branch_0
# incep2_4_0 = ad.convolution_2d_forward_op(concat2_3, filter2_4_0, "NCHW", "SAME", 1, 1)
# # branch 1
# incep2_4_1a = ad.convolution_2d_forward_op(concat2_3, filter2_4_1a, "NCHW", "SAME", 1, 1)
# incep2_4_1b = ad.convolution_2d_forward_op(incep2_4_1a, filter2_4_1b, "NCHW", "SAME", 1, 1)
# incep2_4_1 = ad.convolution_2d_forward_op(incep2_4_1b, filter2_4_1c, "NCHW", "SAME", 1, 1)
# # branch 2
# incep2_4_2a = ad.convolution_2d_forward_op(concat2_3, filter2_4_2a, "NCHW", "SAME", 1, 1)
# incep2_4_2b = ad.convolution_2d_forward_op(incep2_4_2a, filter2_4_2b, "NCHW", "SAME", 1, 1)
# incep2_4_2c = ad.convolution_2d_forward_op(incep2_4_2b, filter2_4_2c, "NCHW", "SAME", 1, 1)
# incep2_4_2d = ad.convolution_2d_forward_op(incep2_4_2c, filter2_4_2d, "NCHW", "SAME", 1, 1)
# incep2_4_2 = ad.convolution_2d_forward_op(incep2_4_2d, filter2_4_2e, "NCHW", "SAME", 1, 1)
# # branch 3
# incep2_4_3a = ad.pooling_2d_forward_op(concat2_3, "NCHW", "mean", 1, 1, 1, 1, 3, 3)
# incep2_4_3 = ad.convolution_2d_forward_op(incep2_4_3a, filter2_4_3, "NCHW", "SAME", 1, 1)
#
# concat2_4a = ad.concat_forward_op(incep2_4_0, incep2_4_1)
# concat2_4b = ad.concat_forward_op(concat2_4a, incep2_4_2)
# concat2_4 = ad.concat_forward_op(concat2_4b, incep2_4_3)
# inception_moudle2_5
# filter2_5_0 = ad.Variable("filter2_5_0")
# filter2_5_1a = ad.Variable("filter2_5_1a")
# filter2_5_1b = ad.Variable("filter2_5_1b")
# filter2_5_1c = ad.Variable("filter2_5_1c")
# filter2_5_2a = ad.Variable("filter2_5_2a")
# filter2_5_2b = ad.Variable("filter2_5_2b")
# filter2_5_2c = ad.Variable("filter2_5_2c")
# filter2_5_2d = ad.Variable("filter2_5_2d")
# filter2_5_2e = ad.Variable("filter2_5_2e")
# filter2_5_3 = ad.Variable("filter2_5_3a")
#
# filter2_5_0_val = np.ones((192, 768, 1, 1)) * 0.01
# filter2_5_1_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_5_1_valb = np.ones((160, 160, 1, 7)) * 0.01
# filter2_5_1_valc = np.ones((192, 160, 7, 1)) * 0.01
# filter2_5_2_vala = np.ones((160, 768, 1, 1)) * 0.01
# filter2_5_2_valb = np.ones((160, 160, 7, 1)) * 0.01
# filter2_5_2_valc = np.ones((160, 160, 1, 7)) * 0.01
# filter2_5_2_vald = np.ones((160, 160, 7, 1)) * 0.01
# filter2_5_2_vale = np.ones((192, 160, 1, 7)) * 0.01
# filter2_5_3_val = np.ones((192, 768, 1, 1)) * 0.01
#
# # branch_0
# incep2_5_0 = ad.convolution_2d_forward_op(concat2_4, filter2_5_0, "NCHW", "SAME", 1, 1)
# # branch 1
# incep2_5_1a = ad.convolution_2d_forward_op(concat2_4, filter2_5_1a, "NCHW", "SAME", 1, 1)
# incep2_5_1b = ad.convolution_2d_forward_op(incep2_5_1a, filter2_5_1b, "NCHW", "SAME", 1, 1)
# incep2_5_1 = ad.convolution_2d_forward_op(incep2_5_1b, filter2_5_1c, "NCHW", "SAME", 1, 1)
# # branch 2
# incep2_5_2a = ad.convolution_2d_forward_op(concat2_4, filter2_5_2a, "NCHW", "SAME", 1, 1)
# incep2_5_2b = ad.convolution_2d_forward_op(incep2_5_2a, filter2_5_2b, "NCHW", "SAME", 1, 1)
# incep2_5_2c = ad.convolution_2d_forward_op(incep2_5_2b, filter2_5_2c, "NCHW", "SAME", 1, 1)
# incep2_5_2d = ad.convolution_2d_forward_op(incep2_5_2c, filter2_5_2d, "NCHW", "SAME", 1, 1)
# incep2_5_2 = ad.convolution_2d_forward_op(incep2_5_2d, filter2_5_2e, "NCHW", "SAME", 1, 1)
# # branch 3
# incep2_5_3a = ad.pooling_2d_forward_op(concat2_4, "NCHW", "mean", 1, 1, 1, 1, 3, 3)
# incep2_5_3 = ad.convolution_2d_forward_op(incep2_5_3a, filter2_5_3, "NCHW", "SAME", 1, 1)
#
# concat2_5a = ad.concat_forward_op(incep2_5_0, incep2_5_1)
# concat2_5b = ad.concat_forward_op(concat2_5a, incep2_5_2)
# concat2_5 = ad.concat_forward_op(concat2_5b, incep2_5_3)
# # inception_moudle3
# inception_moudle3_1
filter3_1_0a = ad.Variable("filter3_1_0a")
filter3_1_0b = ad.Variable("filter3_1_0b")
filter3_1_1a = ad.Variable("filter3_1_1a")
filter3_1_1b = ad.Variable("filter3_1_1b")
filter3_1_1c = ad.Variable("filter3_1_1c")
filter3_1_1d = ad.Variable("filter3_1_1d")
filter3_1_0_vala = np.ones((192, 768, 1, 1)) * 0.01
filter3_1_0_valb = np.ones((320, 192, 3, 3)) * 0.01
filter3_1_1_vala = np.ones((192, 768, 1, 1)) * 0.01
filter3_1_1_valb = np.ones((192, 192, 1, 7)) * 0.01
filter3_1_1_valc = np.ones((192, 192, 7, 1)) * 0.01
filter3_1_1_vald = np.ones((192, 192, 3, 3)) * 0.01
# branch_0
incep3_1_0a = ad.convolution_2d_forward_op(concat2_2, filter3_1_0a, "NCHW",
"SAME", 1, 1)
incep3_1_0 = ad.convolution_2d_forward_op(incep3_1_0a, filter3_1_0b,
"NCHW", "VALID", 2, 2)
# branch 1
incep3_1_1a = ad.convolution_2d_forward_op(concat2_2, filter3_1_1a, "NCHW",
"SAME", 1, 1)
incep3_1_1b = ad.convolution_2d_forward_op(incep3_1_1a, filter3_1_1b,
"NCHW", "SAME", 1, 1)
incep3_1_1c = ad.convolution_2d_forward_op(incep3_1_1b, filter3_1_1c,
"NCHW", "SAME", 1, 1)
incep3_1_1 = ad.convolution_2d_forward_op(incep3_1_1c, filter3_1_1d,
"NCHW", "VALID", 2, 2)
# branch 2
incep3_1_2 = ad.pooling_2d_forward_op(concat2_2, "NCHW", "mean", 0, 0, 2,
2, 3, 3)
concat3_1a = ad.concat_forward_op(incep3_1_0, incep3_1_1)
concat3_1 = ad.concat_forward_op(concat3_1a, incep3_1_2)
# inception_moudle3_2
filter3_2_0 = ad.Variable("filter3_2_0")
filter3_2_1a = ad.Variable("filter3_2_1a")
filter3_2_1b = ad.Variable("filter3_2_1b")
filter3_2_1c = ad.Variable("filter3_2_1c")
filter3_2_2a = ad.Variable("filter3_2_2a")
filter3_2_2b = ad.Variable("filter3_2_2b")
filter3_2_2c = ad.Variable("filter3_2_2c")
filter3_2_2d = ad.Variable("filter3_2_2d")
filter3_2_3 = ad.Variable("filter3_2_3a")
filter3_2_0_val = np.ones((320, 1280, 1, 1)) * 0.01
filter3_2_1_vala = np.ones((384, 1280, 1, 1)) * 0.01
filter3_2_1_valb = np.ones((384, 384, 1, 3)) * 0.01
filter3_2_1_valc = np.ones((384, 384, 3, 1)) * 0.01
filter3_2_2_vala = np.ones((448, 1280, 1, 1)) * 0.01
filter3_2_2_valb = np.ones((384, 448, 3, 3)) * 0.01
filter3_2_2_valc = np.ones((384, 384, 1, 3)) * 0.01
filter3_2_2_vald = np.ones((384, 384, 3, 1)) * 0.01
filter3_2_3_val = np.ones((192, 1280, 1, 1)) * 0.01
# branch_0
incep3_2_0 = ad.convolution_2d_forward_op(concat3_1, filter3_2_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep3_2_1a = ad.convolution_2d_forward_op(concat3_1, filter3_2_1a, "NCHW",
"SAME", 1, 1)
incep3_2_1b = ad.convolution_2d_forward_op(incep3_2_1a, filter3_2_1b,
"NCHW", "SAME", 1, 1)
incep3_2_1c = ad.convolution_2d_forward_op(incep3_2_1a, filter3_2_1c,
"NCHW", "SAME", 1, 1)
incep3_2_1 = ad.concat_forward_op(incep3_2_1b, incep3_2_1c)
# branch 2
incep3_2_2a = ad.convolution_2d_forward_op(concat3_1, filter3_2_2a, "NCHW",
"SAME", 1, 1)
incep3_2_2b = ad.convolution_2d_forward_op(incep3_2_2a, filter3_2_2b,
"NCHW", "SAME", 1, 1)
incep3_2_2c = ad.convolution_2d_forward_op(incep3_2_2b, filter3_2_2c,
"NCHW", "SAME", 1, 1)
incep3_2_2d = ad.convolution_2d_forward_op(incep3_2_2b, filter3_2_2d,
"NCHW", "SAME", 1, 1)
incep3_2_2 = ad.concat_forward_op(incep3_2_2c, incep3_2_2d)
# branch 3
incep3_2_3a = ad.pooling_2d_forward_op(concat3_1, "NCHW", "mean", 1, 1, 1,
1, 3, 3)
incep3_2_3 = ad.convolution_2d_forward_op(incep3_2_3a, filter3_2_3, "NCHW",
"SAME", 1, 1)
concat3_2a = ad.concat_forward_op(incep3_2_0, incep3_2_1)
concat3_2b = ad.concat_forward_op(concat3_2a, incep3_2_2)
concat3_2 = ad.concat_forward_op(concat3_2b, incep3_2_3)
# # inception_moudle3_3
# filter3_3_0 = ad.Variable("filter3_3_0")
# filter3_3_1a = ad.Variable("filter3_3_1a")
# filter3_3_1b = ad.Variable("filter3_3_1b")
# filter3_3_1c = ad.Variable("filter3_3_1c")
# filter3_3_2a = ad.Variable("filter3_3_2a")
# filter3_3_2b = ad.Variable("filter3_3_2b")
# filter3_3_2c = ad.Variable("filter3_3_2c")
# filter3_3_2d = ad.Variable("filter3_3_2d")
# filter3_3_3 = ad.Variable("filter3_3_3a")
#
# filter3_3_0_val = np.ones((320, 2048, 1, 1)) * 0.01
# filter3_3_1_vala = np.ones((384, 2048, 1, 1)) * 0.01
# filter3_3_1_valb = np.ones((384, 384, 1, 3)) * 0.01
# filter3_3_1_valc = np.ones((384, 384, 3, 1)) * 0.01
# filter3_3_2_vala = np.ones((448, 2048, 1, 1)) * 0.01
# filter3_3_2_valb = np.ones((384, 448, 3, 3)) * 0.01
# filter3_3_2_valc = np.ones((384, 384, 1, 3)) * 0.01
# filter3_3_2_vald = np.ones((384, 384, 3, 1)) * 0.01
# filter3_3_3_val = np.ones((192, 2048, 1, 1)) * 0.01
#
# # branch_0
# incep3_3_0 = ad.convolution_2d_forward_op(concat3_2, filter3_3_0, "NCHW", "SAME", 1, 1)
# # branch 1
# incep3_3_1a = ad.convolution_2d_forward_op(concat3_2, filter3_3_1a, "NCHW", "SAME", 1, 1)
# incep3_3_1b = ad.convolution_2d_forward_op(incep3_3_1a, filter3_3_1b, "NCHW", "SAME", 1, 1)
# incep3_3_1c = ad.convolution_2d_forward_op(incep3_3_1a, filter3_3_1c, "NCHW", "SAME", 1, 1)
# incep3_3_1 = ad.concat_forward_op(incep3_3_1b, incep3_3_1c)
#
# # branch 2
# incep3_3_2a = ad.convolution_2d_forward_op(concat3_2, filter3_3_2a, "NCHW", "SAME", 1, 1)
# incep3_3_2b = ad.convolution_2d_forward_op(incep3_3_2a, filter3_3_2b, "NCHW", "SAME", 1, 1)
# incep3_3_2c = ad.convolution_2d_forward_op(incep3_3_2b, filter3_3_2c, "NCHW", "SAME", 1, 1)
# incep3_3_2d = ad.convolution_2d_forward_op(incep3_3_2b, filter3_3_2d, "NCHW", "SAME", 1, 1)
# incep3_3_2 = ad.concat_forward_op(incep3_3_2c, incep3_3_2d)
# # branch 3
# incep3_3_3a = ad.pooling_2d_forward_op(concat3_2, "NCHW", "mean", 1, 1, 1, 1, 3, 3)
# incep3_3_3 = ad.convolution_2d_forward_op(incep3_3_3a, filter3_3_3, "NCHW", "SAME", 1, 1)
#
# concat3_3a = ad.concat_forward_op(incep3_3_0, incep3_3_1)
# concat3_3b = ad.concat_forward_op(concat3_3a, incep3_3_2)
# concat3_3 = ad.concat_forward_op(concat3_3b, incep3_3_3)
filtera1 = ad.Variable("filtera1")
filtera1val = np.ones((1000, 2048, 1, 1)) * 0.01
filtersmul = ad.Variable("filtersmul")
filtersmulval = np.ones((1000, n_class)) * 0.01
biases = ad.Variable("biases")
biasesval = np.ones((batch_size, n_class)) * 0.01
poollast = ad.pooling_2d_forward_op(concat3_2, "NCHW", "mean", 0, 0, 1, 1,
8, 8)
dropout = ad.dropout_forward_op(poollast, "NCHW", 0.8)
convlast = ad.convolution_2d_forward_op(dropout, filtera1, "NCHW", "SAME",
1, 1)
squeeze = ad.squeeze_op(convlast)
# fc8
mul8 = ad.matmul_op(squeeze, filtersmul)
add8 = ad.add_op(mul8, biases)
fc8 = ad.fullyactivation_forward_op(add8, "NCHW", "softmax")
loss = ad.softmaxcrossentropy_op(fc8, y_)
X_val = np.empty(shape=(batch_size, 3, 32, 32), dtype=np.float32)
y_val = np.empty(shape=(batch_size, n_class), dtype=np.float32)
aph = 0.001
t = train.Adam_minimize(loss, aph)
t.init_Variable({
filterb_1: filtersb_val1,
filterb_2: filtersb_val2,
filterb_3: filtersb_val3,
filterb_4: filtersb_val4,
filterb_5: filtersb_val5,
filter1_1_0: filter1_1_0_val,
filter1_1_1a: filter1_1_1_vala,
filter1_1_1b: filter1_1_1_valb,
filter1_1_2a: filter1_1_2_vala,
filter1_1_2b: filter1_1_2_valb,
filter1_1_2c: filter1_1_2_valc,
filter1_1_3: filter1_1_3_val,
filter1_2_0: filter1_2_0_val,
filter1_2_1a: filter1_2_1_vala,
filter1_2_1b: filter1_2_1_valb,
filter1_2_2a: filter1_2_2_vala,
filter1_2_2b: filter1_2_2_valb,
filter1_2_2c: filter1_2_2_valc,
filter1_2_3: filter1_2_3_val,
filter1_3_0: filter1_3_0_val,
filter1_3_1a: filter1_3_1_vala,
filter1_3_1b: filter1_3_1_valb,
filter1_3_2a: filter1_3_2_vala,
filter1_3_2b: filter1_3_2_valb,
filter1_3_2c: filter1_3_2_valc,
filter1_3_3: filter1_3_3_val,
filter2_1_0: filter2_1_0_val,
filter2_1_1a: filter2_1_1_vala,
filter2_1_1b: filter2_1_1_valb,
filter2_1_1c: filter2_1_1_valc,
filter2_2_0: filter2_2_0_val,
filter2_2_1a: filter2_2_1_vala,
filter2_2_1b: filter2_2_1_valb,
filter2_2_1c: filter2_2_1_valc,
filter2_2_2a: filter2_2_2_vala,
filter2_2_2b: filter2_2_2_valb,
filter2_2_2c: filter2_2_2_valc,
filter2_2_2d: filter2_2_2_vald,
filter2_2_2e: filter2_2_2_vale,
filter2_2_3: filter2_2_3_val
# , filter2_3_0: filter2_3_0_val, filter2_3_1a: filter2_3_1_vala, filter2_3_1b: filter2_3_1_valb,
# filter2_3_1c: filter2_3_1_valc,
# filter2_3_2a: filter2_3_2_vala, filter2_3_2b: filter2_3_2_valb,
# filter2_3_2c: filter2_3_2_valc, filter2_3_2d: filter2_3_2_vald,
# filter2_3_2e: filter2_3_2_vale, filter2_3_3: filter2_3_3_val
# , filter2_4_0: filter2_4_0_val, filter2_4_1a: filter2_4_1_vala, filter2_4_1b: filter2_4_1_valb,
# filter2_4_1c: filter2_4_1_valc,
# filter2_4_2a: filter2_4_2_vala, filter2_4_2b: filter2_4_2_valb,
# filter2_4_2c: filter2_4_2_valc, filter2_4_2d: filter2_4_2_vald,
# filter2_4_2e: filter2_4_2_vale, filter2_4_3: filter2_4_3_val
# , filter2_5_0: filter2_5_0_val, filter2_5_1a: filter2_5_1_vala, filter2_5_1b: filter2_5_1_valb,
# filter2_5_1c: filter2_5_1_valc,
# filter2_5_2a: filter2_5_2_vala, filter2_5_2b: filter2_5_2_valb,
# filter2_5_2c: filter2_5_2_valc, filter2_5_2d: filter2_5_2_vald,
# filter2_5_2e: filter2_5_2_vale, filter2_5_3: filter2_5_3_val
,
filter3_1_0a: filter3_1_0_vala,
filter3_1_0b: filter3_1_0_valb,
filter3_1_1a: filter3_1_1_vala,
filter3_1_1b: filter3_1_1_valb,
filter3_1_1c: filter3_1_1_valc,
filter3_1_1d: filter3_1_1_vald,
filter3_2_0: filter3_2_0_val,
filter3_2_1a: filter3_2_1_vala,
filter3_2_1b: filter3_2_1_valb,
filter3_2_1c: filter3_2_1_valc,
filter3_2_2a: filter3_2_2_vala,
filter3_2_2b: filter3_2_2_valb,
filter3_2_2c: filter3_2_2_valc,
filter3_2_2d: filter3_2_2_vald,
filter3_2_3: filter3_2_3_val
# , filter3_3_0: filter3_3_0_val, filter3_3_1a: filter3_3_1_vala,
# filter3_3_1b: filter3_3_1_valb,
# filter3_3_1c: filter3_3_1_valc, filter3_3_2a: filter3_3_2_vala,
# filter3_3_2b: filter3_3_2_valb,
# filter3_3_2c: filter3_3_2_valc, filter3_3_2d: filter3_3_2_vald,
# filter3_3_3: filter3_3_3_val,
,
filtera1: filtera1val,
filtersmul: filtersmulval,
biases: biasesval
})
train_x, train_y, test_x, test_y = prepare_data()
num_epochs = 10
n_train_batches = train_x.shape[0] // batch_size
n_test_batches = test_x.shape[0] // batch_size
for i in range(num_epochs):
print("epoch %d" % i)
for minibatch_index in range(n_train_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_val[:] = train_x[minibatch_start:minibatch_end]
y_val[:] = train_y[minibatch_start:minibatch_end]
# print(y_val.shape)
# y_val[:] = convert_to_one_hot(train_y[minibatch_start:minibatch_end])
t.run({X: X_val, y_: y_val})
# print(t.run_get_nodelist_once({X: X_val, y_: y_val}, [fc8])[fc8].asnumpy())
# print(map_to_numpy(t.get_Variable_node_to_val_map()))
print("loss_val:", t.get_loss().asnumpy())
def test_inception_v4():
X = ad.Placeholder("X")
Y_ = ad.Placeholder("y_")
f1 = ad.Variable("f1")
f2 = ad.Variable("f2")
f3 = ad.Variable("f3")
f4 = ad.Variable("f4")
f5_1 = ad.Variable("f5_1")
f5_2 = ad.Variable("f5_2")
f6_1 = ad.Variable("f6_1")
f6_2 = ad.Variable("f6_2")
f6_3 = ad.Variable("f6_3")
f6_4 = ad.Variable("f6_4")
f7 = ad.Variable("f7")
W = ad.Variable("W")
b = ad.Variable("b")
keep_prob = ad.Placeholder("keep_prob")
X_val = np.random.normal(0, 0.5, (10, 3, 299, 299))
Y_val = np.random.normal(0, 0.5, (10, 1))
f1val = np.random.normal(0, 0.5, (32, 3, 3, 3))
f2val = np.random.normal(0, 0.5, (32, 32, 3, 3))
f3val = np.random.normal(0, 0.5, (64, 32, 3, 3))
f4val = np.random.normal(0, 0.5, (96, 64, 3, 3))
f5_1val = np.random.normal(0, 0.5, (64, 160, 1, 1))
f5_2val = np.random.normal(0, 0.5, (96, 64, 3, 3))
f6_1val = np.random.normal(0, 0.5, (64, 160, 1, 1))
f6_2val = np.random.normal(0, 0.5, (64, 64, 7, 1))
f6_3val = np.random.normal(0, 0.5, (64, 64, 1, 7))
f6_4val = np.random.normal(0, 0.5, (96, 64, 3, 3))
f7val = np.random.normal(0, 0.5, (192, 192, 3, 3))
# stem
cov1 = ad.convolution_2d_forward_op(X, f1, "NCHW", "VALID", 2, 2)
cov2 = ad.convolution_2d_forward_op(cov1, f2, "NCHW", "VALID", 1, 1)
cov3 = ad.convolution_2d_forward_op(cov2, f3, "NCHW", "SAME", 1, 1)
pool4 = ad.pooling_2d_forward_op(cov3, "NCHW", "max", 0, 0, 2, 2, 3, 3)
cov4 = ad.convolution_2d_forward_op(cov3, f4, "NCHW", "VALID", 2, 2)
concat1 = ad.concat_forward_op(pool4, cov4)
cov5_1 = ad.convolution_2d_forward_op(concat1, f5_1, "NCHW", "SAME", 1, 1)
cov5_2 = ad.convolution_2d_forward_op(cov5_1, f5_2, "NCHW", "VALID", 1, 1)
cov6_1 = ad.convolution_2d_forward_op(concat1, f6_1, "NCHW", "SAME", 1, 1)
cov6_2 = ad.convolution_2d_forward_op(cov6_1, f6_2, "NCHW", "SAME", 1, 1)
cov6_3 = ad.convolution_2d_forward_op(cov6_2, f6_3, "NCHW", "SAME", 1, 1)
cov6_4 = ad.convolution_2d_forward_op(cov6_3, f6_4, "NCHW", "VALID", 1, 1)
concat2 = ad.concat_forward_op(cov5_2, cov6_4)
cov7 = ad.convolution_2d_forward_op(concat2, f7, "NCHW", "VALID", 2, 2)
pool7 = ad.pooling_2d_forward_op(concat2, "NCHW", "max", 0, 0, 2, 2, 3, 3)
concat3 = ad.concat_forward_op(pool7, cov7)
a1, dicta1 = block_inception_a(concat3, "a1")
a2, dicta2 = block_inception_a(a1, "a2")
a3, dicta3 = block_inception_a(a2, "a3")
a4, dicta4 = block_inception_a(a3, "a4")
ra, dictra = block_reduction_a(a4, 384, "ra")
b1, dictb1 = block_inception_b(ra, 1024, "b1")
# b2, dictb2 = block_inception_b(b1, 1024, "b2")
# b3, dictb3 = block_inception_b(b2, 1024, "b3")
# b4, dictb4 = block_inception_b(b3, 1024, "b4")
# b5, dictb5 = block_inception_b(b4, 1024, "b5")
# b6, dictb6 = block_inception_b(b5, 1024, "b6")
# b7, dictb7 = block_inception_b(b6, 1024, "b7")
#
rb, dictrb = block_reduction_b(b1, 1024, "rb")
c1, dictc1 = block_inception_c(rb, 1536, "c1")
# c2, dictc2 = block_inception_c(c1, 1536, "c2")
# c3, dictc3 = block_inception_c(c2, 1536, "c3")
poollast = ad.pooling_2d_forward_op(c1, "NCHW", "mean", 0, 0, 1, 1, 8, 8)
squeeze = ad.squeeze_op(poollast)
drop_out = ad.fullydropout_forward_op(squeeze, "NCHW", 0.8)
mul = ad.matmul_op(drop_out, W)
add = ad.add_op(mul, b)
loss = ad.softmaxcrossentropy_op(add, Y_)
#
# aph = 0.001
# t = train.Adam_minimize(loss, aph)
ctx = ndarray.gpu(0)
# aph = 0.001
# t = train.Adam_minimize(loss, aph)
feed_dict = {
X: X_val,
f1: f1val,
f2: f2val,
f3: f3val,
f4: f4val,
f5_1: f5_1val,
f5_2: f5_2val,
f6_1: f6_1val,
f6_2: f6_2val,
f6_3: f6_3val,
f6_4: f6_4val,
f7: f7val
}
feed_dict.update(dicta1)
feed_dict.update(dicta2)
feed_dict.update(dicta3)
feed_dict.update(dicta4)
feed_dict.update(dictra)
feed_dict.update(dictb1)
# feed_dict.update(dictb2)
# feed_dict.update(dictb3)
# feed_dict.update(dictb4)
# feed_dict.update(dictb5)
# feed_dict.update(dictb6)
# feed_dict.update(dictb7)
feed_dict.update(dictrb)
feed_dict.update(dictc1)
# feed_dict.update(dictc2)
# feed_dict.update(dictc3)
executor = ad.Executor([drop_out], ctx=ctx)
y_val = executor.run(feed_dict)
def block_inception_c(inputs, input_size, blockname=None):
filter3_2_0 = ad.Variable(blockname + "filter3_2_0")
filter3_2_1a = ad.Variable(blockname + "filter3_2_1a")
filter3_2_1b = ad.Variable(blockname + "filter3_2_1b")
filter3_2_1c = ad.Variable(blockname + "filter3_2_1c")
filter3_2_2a = ad.Variable(blockname + "filter3_2_2a")
filter3_2_2b = ad.Variable(blockname + "filter3_2_2b")
filter3_2_2c = ad.Variable(blockname + "filter3_2_2c")
filter3_2_2d = ad.Variable(blockname + "filter3_2_2d")
filter3_2_2e = ad.Variable(blockname + "filter3_2_2e")
filter3_2_3 = ad.Variable(blockname + "filter3_2_3a")
rand = np.random.RandomState(seed=123)
filter3_2_0_val = rand.normal(scale=0.1, size=(256, input_size, 1, 1))
filter3_2_1_vala = rand.normal(scale=0.1, size=(384, input_size, 1, 1))
filter3_2_1_valb = rand.normal(scale=0.1, size=(256, 384, 1, 3))
filter3_2_1_valc = rand.normal(scale=0.1, size=(256, 384, 3, 1))
filter3_2_2_vala = rand.normal(scale=0.1, size=(384, input_size, 1, 1))
filter3_2_2_valb = rand.normal(scale=0.1, size=(448, 384, 1, 3))
filter3_2_2_valc = rand.normal(scale=0.1, size=(512, 448, 3, 1))
filter3_2_2_vald = rand.normal(scale=0.1, size=(256, 512, 3, 1))
filter3_2_2_vale = rand.normal(scale=0.1, size=(256, 512, 1, 3))
filter3_2_3_val = rand.normal(scale=0.1, size=(256, input_size, 1, 1))
# branch_0
incep3_2_0 = ad.convolution_2d_forward_op(inputs, filter3_2_0, "NCHW",
"SAME", 1, 1)
# branch 1
incep3_2_1a = ad.convolution_2d_forward_op(inputs, filter3_2_1a, "NCHW",
"SAME", 1, 1)
incep3_2_1b = ad.convolution_2d_forward_op(incep3_2_1a, filter3_2_1b,
"NCHW", "SAME", 1, 1)
incep3_2_1c = ad.convolution_2d_forward_op(incep3_2_1a, filter3_2_1c,
"NCHW", "SAME", 1, 1)
incep3_2_1 = ad.concat_forward_op(incep3_2_1b, incep3_2_1c)
# branch 2
incep3_2_2a = ad.convolution_2d_forward_op(inputs, filter3_2_2a, "NCHW",
"SAME", 1, 1)
incep3_2_2b = ad.convolution_2d_forward_op(incep3_2_2a, filter3_2_2b,
"NCHW", "SAME", 1, 1)
incep3_2_2c = ad.convolution_2d_forward_op(incep3_2_2b, filter3_2_2c,
"NCHW", "SAME", 1, 1)
incep3_2_2d = ad.convolution_2d_forward_op(incep3_2_2c, filter3_2_2d,
"NCHW", "SAME", 1, 1)
incep3_2_2e = ad.convolution_2d_forward_op(incep3_2_2c, filter3_2_2e,
"NCHW", "SAME", 1, 1)
incep3_2_2 = ad.concat_forward_op(incep3_2_2d, incep3_2_2e)
# branch 3
incep3_2_3a = ad.pooling_2d_forward_op(inputs, "NCHW", "mean", 1, 1, 1, 1,
3, 3)
incep3_2_3 = ad.convolution_2d_forward_op(incep3_2_3a, filter3_2_3, "NCHW",
"SAME", 1, 1)
concat3_2a = ad.concat_forward_op(incep3_2_0, incep3_2_1)
concat3_2b = ad.concat_forward_op(concat3_2a, incep3_2_2)
concat3_2 = ad.concat_forward_op(concat3_2b, incep3_2_3)
dict = {
filter3_2_0: filter3_2_0_val,
filter3_2_1a: filter3_2_1_vala,
filter3_2_1b: filter3_2_1_valb,
filter3_2_1c: filter3_2_1_valc,
filter3_2_2a: filter3_2_2_vala,
filter3_2_2b: filter3_2_2_valb,
filter3_2_2c: filter3_2_2_valc,
filter3_2_2d: filter3_2_2_vald,
filter3_2_2e: filter3_2_2_vale,
filter3_2_3: filter3_2_3_val
}
return concat3_2, dict
def mnist_logreg(executor_ctx=None, num_epochs=10, print_loss_val_each_epoch=False):
# 训练逻辑回归模型
print("Build logistic regression model...")
W1 = ad.Variable(name="W1")
b1 = ad.Variable(name="b1")
X = ad.Variable(name="X")
y_ = ad.Variable(name="y_")
z1 = ad.matmul_op(X, W1)
y = z1 + ad.broadcastto_op(b1, z1)
loss = ad.softmaxcrossentropy_op(y, y_)
grad_W1, grad_b1 = ad.gradients(loss, [W1, b1])
executor = ad.Executor([loss, grad_W1, grad_b1, y], ctx=executor_ctx)
# Read input data
datasets = load_mnist_data("mnist.pkl.gz")
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# Set up minibatch
batch_size = 1000
n_train_batches = train_set_x.shape[0] // batch_size
n_valid_batches = valid_set_x.shape[0] // batch_size
print("Start training loop...")
# Initialize parameters
W1_val = np.zeros((784, 10))
b1_val = np.zeros((10))
X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
valid_X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
valid_y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
if ndarray.is_gpu_ctx(executor_ctx):
W1_val = ndarray.array(W1_val, ctx=executor_ctx)
b1_val = ndarray.array(b1_val, ctx=executor_ctx)
X_val = ndarray.array(X_val, ctx=executor_ctx)
y_val = ndarray.array(y_val, ctx=executor_ctx)
lr = 1e-3
for i in range(num_epochs):
print("epoch %d" % i)
for minibatch_index in range(n_train_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_val[:] = train_set_x[minibatch_start:minibatch_end]
y_val[:] = convert_to_one_hot(
train_set_y[minibatch_start:minibatch_end])
loss_val, grad_W1_val, grad_b1_val, _ = executor.run(
feed_dict = {X: X_val, y_: y_val, W1: W1_val, b1: b1_val})
# SGD update
if (executor_ctx is None):
W1_val = W1_val - lr * grad_W1_val
b1_val = b1_val - lr * grad_b1_val
else:
sgd_update_gpu(W1_val, grad_W1_val, lr)
sgd_update_gpu(b1_val, grad_b1_val, lr)
if print_loss_val_each_epoch:
if isinstance(loss_val, ndarray.NDArray):
print(loss_val.asnumpy())
else:
print(loss_val)
correct_predictions = []
for minibatch_index in range(n_valid_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
valid_X_val[:] = valid_set_x[minibatch_start:minibatch_end]
valid_y_val[:] = convert_to_one_hot(
valid_set_y[minibatch_start:minibatch_end])
_, _, _, valid_y_predicted = executor.run(
feed_dict={
X: valid_X_val,
y_: valid_y_val,
W1: W1_val,
b1: b1_val},
convert_to_numpy_ret_vals=True)
correct_prediction = np.equal(
np.argmax(valid_y_val, 1),
np.argmax(valid_y_predicted, 1)).astype(np.float)
correct_predictions.extend(correct_prediction)
accuracy = np.mean(correct_predictions)
# validation set accuracy=0.928200
print("validation set accuracy=%f" % accuracy)
def mnist_mlp(executor_ctx=None, num_epochs=10, print_loss_val_each_epoch=False):
# 训练一个三层感知机模型
print("Build 3-layer MLP model...")
W1 = ad.Variable(name="W1")
W2 = ad.Variable(name="W2")
W3 = ad.Variable(name="W3")
b1 = ad.Variable(name="b1")
b2 = ad.Variable(name="b2")
b3 = ad.Variable(name="b3")
X = ad.Variable(name="X")
y_ = ad.Variable(name="y_")
# 下面是三层网络的激活函数，两个relu和一个softmax
# relu(X W1+b1)
z1 = ad.matmul_op(X, W1)
z2 = z1 + ad.broadcastto_op(b1, z1)
z3 = ad.relu_op(z2)
# relu(z3 W2+b2)
z4 = ad.matmul_op(z3, W2)
z5 = z4 + ad.broadcastto_op(b2, z4)
z6 = ad.relu_op(z5)
# softmax(z5 W2+b2)
z7 = ad.matmul_op(z6, W3)
y = z7 + ad.broadcastto_op(b3, z7)
loss = ad.softmaxcrossentropy_op(y, y_)
grad_W1, grad_W2, grad_W3, grad_b1, grad_b2, grad_b3 = ad.gradients(
loss, [W1, W2, W3, b1, b2, b3])
# 此处向前为符号定义
# 只声明，不操作
executor = ad.Executor(
[loss, grad_W1, grad_W2, grad_W3, grad_b1, grad_b2, grad_b3, y],
ctx=executor_ctx)
# Read input data
datasets = load_mnist_data("mnist.pkl.gz")
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# Set up minibatch
batch_size = 1000
n_train_batches = train_set_x.shape[0] // batch_size
n_valid_batches = valid_set_x.shape[0] // batch_size
print("Start training loop...")
# Initialize parameters
# 随机初始化网络中的w和b
rand = np.random.RandomState(seed=123)
W1_val = rand.normal(scale=0.1, size=(784, 256))
W2_val = rand.normal(scale=0.1, size=(256, 100))
W3_val = rand.normal(scale=0.1, size=(100, 10))
b1_val = rand.normal(scale=0.1, size=(256))
b2_val = rand.normal(scale=0.1, size=(100))
b3_val = rand.normal(scale=0.1, size=(10))
X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
valid_X_val = np.empty(shape=(batch_size, 784), dtype=np.float32)
valid_y_val = np.empty(shape=(batch_size, 10), dtype=np.float32)
# todo 此处修改为cpu
W1_val = ndarray.array(W1_val, ctx=executor_ctx_cpu)
W2_val = ndarray.array(W2_val, ctx=executor_ctx_cpu)
W3_val = ndarray.array(W3_val, ctx=executor_ctx_cpu)
b1_val = ndarray.array(b1_val, ctx=executor_ctx_cpu)
b2_val = ndarray.array(b2_val, ctx=executor_ctx_cpu)
b3_val = ndarray.array(b3_val, ctx=executor_ctx_cpu)
X_val = ndarray.array(X_val, ctx=executor_ctx_cpu)
y_val = ndarray.array(y_val, ctx=executor_ctx_cpu)
# 此处以上将数据分别转化为cpu和gpu两种格式
lr = 1.0e-3
for i in range(num_epochs):
print("epoch %d" % i)
for minibatch_index in range(n_train_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_val[:] = train_set_x[minibatch_start:minibatch_end]
y_val[:] = convert_to_one_hot(
train_set_y[minibatch_start:minibatch_end])
# 计算单步的梯度
loss_val, grad_W1_val, grad_W2_val, grad_W3_val, \
grad_b1_val, grad_b2_val, grad_b3_val, _ = executor.run(
feed_dict={
X: X_val,
y_: y_val,
W1: W1_val,
W2: W2_val,
W3: W3_val,
b1: b1_val,
b2: b2_val,
b3: b3_val})
# todo 更新sgd_update_gpu_on_cpu
def sgd_update_cpu(w1, w2, w3):
w1_gpu = ndarray.empty(w1.shape, executor_ctx)
w1.copyto(w1_gpu)
w2_gpu = ndarray.empty(w2.shape, executor_ctx)
w2.copyto(w2_gpu)
sgd_update_gpu(w1_gpu, w2_gpu, w3)
w1_gpu.copyto(w1)
w2_gpu.copyto(w2)
sgd_update_cpu(W1_val, grad_W1_val, lr)
sgd_update_cpu(W2_val, grad_W2_val, lr)
sgd_update_cpu(W3_val, grad_W3_val, lr)
sgd_update_cpu(b1_val, grad_b1_val, lr)
sgd_update_cpu(b2_val, grad_b2_val, lr)
sgd_update_cpu(b3_val, grad_b3_val, lr)
# sgd_update_gpu(W1_val, grad_W1_val, lr)
# sgd_update_gpu(W2_val, grad_W2_val, lr)
# sgd_update_gpu(W3_val, grad_W3_val, lr)
# sgd_update_gpu(b1_val, grad_b1_val, lr)
# sgd_update_gpu(b2_val, grad_b2_val, lr)
# sgd_update_gpu(b3_val, grad_b3_val, lr)
if print_loss_val_each_epoch:
if isinstance(loss_val, ndarray.NDArray):
print(loss_val.asnumpy())
else:
print(loss_val)
correct_predictions = []
for minibatch_index in range(n_valid_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
valid_X_val[:] = valid_set_x[minibatch_start:minibatch_end]
valid_y_val[:] = convert_to_one_hot(
valid_set_y[minibatch_start:minibatch_end])
_, _, _, _, _, _, _, valid_y_predicted = executor.run(
feed_dict={
X: valid_X_val,
y_: valid_y_val,
W1: W1_val,
W2: W2_val,
W3: W3_val,
b1: b1_val,
b2: b2_val,
b3: b3_val},
convert_to_numpy_ret_vals=True)
correct_prediction = np.equal(
np.argmax(valid_y_val, 1),
np.argmax(valid_y_predicted, 1)).astype(np.float)
correct_predictions.extend(correct_prediction)
accuracy = np.mean(correct_predictions)
# validation set accuracy=0.970800
print("validation set accuracy=%f" % accuracy)
def vgg16():
n = 10
n_class = 10
inputs = ad.Variable("inputs")
filters1_1 = ad.Variable("filters1_1")
filters1_2 = ad.Variable("filters1_2")
filters2_1 = ad.Variable("filters2_1")
filters2_2 = ad.Variable("filters2_2")
filters3_1 = ad.Variable("filters3_1")
filters3_2 = ad.Variable("filters3_2")
filters3_3 = ad.Variable("filters3_3")
filters4_1 = ad.Variable("filters4_1")
filters4_2 = ad.Variable("filters4_2")
filters4_3 = ad.Variable("filters4_3")
filters5_1 = ad.Variable("filters5_1")
filters5_2 = ad.Variable("filters5_2")
filters5_3 = ad.Variable("filters5_3")
filters6 = ad.Variable("filters6")
filters7 = ad.Variable("filters7")
filters8 = ad.Variable("filters8")
biases6 = ad.Variable("biases6")
biases7 = ad.Variable("biases7")
biases8 = ad.Variable("biases8")
y_ = ad.Variable(name="y_")
x_val = np.linspace(0, 0.001, 10*3*224*224).reshape((10, 3, 224, 224))
filters_val = [np.ones((64, 3, 3, 3))*0.001]
filters_val.append(np.ones((64, 64, 3, 3))*0.001)
filters_val.append(np.ones((128, 64, 3, 3))*0.001)
filters_val.append(np.ones((128, 128, 3, 3))*0.001)
filters_val.append(np.ones((256, 128, 3, 3))*0.001)
filters_val.append(np.ones((256, 256, 3, 3))*0.001)
filters_val.append(np.ones((256, 256, 3, 3))*0.001)
filters_val.append(np.ones((512, 256, 3, 3))*0.001)
filters_val.append(np.ones((512, 512, 3, 3))*0.001)
filters_val.append(np.ones((512, 512, 3, 3))*0.001)
filters_val.append(np.ones((512, 512, 3, 3))*0.001)
filters_val.append(np.ones((512, 512, 3, 3))*0.001)
filters_val.append(np.ones((512, 512, 3, 3))*0.001)
filters_val.append(np.ones((512*7*7, 4096)) * 0.001)
filters_val.append(np.ones((4096, 4096)) * 0.001)
filters_val.append(np.ones((4096, n_class)) * 0.001)
biases_val = [np.ones((1, 4096))* 0.001]
biases_val.append(np.ones((1, 4096)) * 0.001)
biases_val.append(np.ones((1, n_class)) * 0.001)
y_val = np.zeros((10, n_class))
ctx = ndarray.gpu(0)
for i in range(16):
filters_val[i] = ndarray.array(filters_val[i], ctx)
# conv 1
conv1_1 = ad.convolution_2d_forward_op(inputs, filters1_1, "NCHW", "SAME", 1, 1)
bn1_1 = ad.bn_forward_op(conv1_1, "NCHW", "pre_activation")
act1_1 = ad.activation_forward_op(bn1_1, "NCHW", "relu")
conv1_2 = ad.convolution_2d_forward_op(act1_1, filters1_2, "NCHW", "SAME", 1, 1)
bn1_2 = ad.bn_forward_op(conv1_2, "NCHW", "pre_activation")
act1_2 = ad.activation_forward_op(bn1_2, "NCHW", "relu")
pool1 = ad.pooling_2d_forward_op(act1_2, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 2
conv2_1 = ad.convolution_2d_forward_op(pool1, filters2_1, "NCHW", "SAME", 1, 1)
bn2_1 = ad.bn_forward_op(conv2_1, "NCHW", "pre_activation")
act2_1 = ad.activation_forward_op(bn2_1, "NCHW", "relu")
conv2_2 = ad.convolution_2d_forward_op(act2_1, filters2_2, "NCHW", "SAME", 1, 1)
bn2_2 = ad.bn_forward_op(conv2_2, "NCHW", "pre_activation")
act2_2 = ad.activation_forward_op(bn2_2, "NCHW", "relu")
pool2 = ad.pooling_2d_forward_op(act2_2, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 3
conv3_1 = ad.convolution_2d_forward_op(pool2, filters3_1, "NCHW", "SAME", 1, 1)
bn3_1 = ad.bn_forward_op(conv3_1, "NCHW", "pre_activation")
act3_1 = ad.activation_forward_op(bn3_1, "NCHW", "relu")
conv3_2 = ad.convolution_2d_forward_op(act3_1, filters3_2, "NCHW", "SAME", 1, 1)
bn3_2 = ad.bn_forward_op(conv3_2, "NCHW", "pre_activation")
act3_2 = ad.activation_forward_op(bn3_2, "NCHW", "relu")
conv3_3 = ad.convolution_2d_forward_op(act3_2, filters3_3, "NCHW", "SAME", 1, 1)
bn3_3 = ad.bn_forward_op(conv3_3, "NCHW", "pre_activation")
act3_3 = ad.activation_forward_op(bn3_3, "NCHW", "relu")
pool3 = ad.pooling_2d_forward_op(act3_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 4
conv4_1 = ad.convolution_2d_forward_op(pool3, filters4_1, "NCHW", "SAME", 1, 1)
bn4_1 = ad.bn_forward_op(conv4_1, "NCHW", "pre_activation")
act4_1 = ad.activation_forward_op(bn4_1, "NCHW", "relu")
conv4_2 = ad.convolution_2d_forward_op(act4_1, filters4_2, "NCHW", "SAME", 1, 1)
bn4_2 = ad.bn_forward_op(conv4_2, "NCHW", "pre_activation")
act4_2 = ad.activation_forward_op(bn4_2, "NCHW", "relu")
conv4_3 = ad.convolution_2d_forward_op(act4_2, filters4_3, "NCHW", "SAME", 1, 1)
bn4_3 = ad.bn_forward_op(conv4_3, "NCHW", "pre_activation")
act4_3 = ad.activation_forward_op(bn4_3, "NCHW", "relu")
pool4 = ad.pooling_2d_forward_op(act4_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 5
conv5_1 = ad.convolution_2d_forward_op(pool4, filters5_1, "NCHW", "SAME", 1, 1)
bn5_1 = ad.bn_forward_op(conv5_1, "NCHW", "pre_activation")
act5_1 = ad.activation_forward_op(bn5_1, "NCHW", "relu")
conv5_2 = ad.convolution_2d_forward_op(act5_1, filters5_2, "NCHW", "SAME", 1, 1)
bn5_2 = ad.bn_forward_op(conv5_2, "NCHW", "pre_activation")
act5_2 = ad.activation_forward_op(bn5_2, "NCHW", "relu")
conv5_3 = ad.convolution_2d_forward_op(act5_2, filters5_3, "NCHW", "SAME", 1, 1)
bn5_3 = ad.bn_forward_op(conv5_3, "NCHW", "pre_activation")
act5_3 = ad.activation_forward_op(bn5_3, "NCHW", "relu")
pool5 = ad.pooling_2d_forward_op(act5_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# fc6
pool5_flat = ad.flatten_op(pool5)
mul6 = ad.matmul_op(pool5_flat, filters6)
add6 = ad.add_op(mul6, biases6)
bn6 = ad.fullybn_forward_op(add6, "NCHW")
fc6 = ad.fullyactivation_forward_op(bn6, "NCHW", "relu")
drop6 = ad.fullydropout_forward_op(fc6, "NCHW", 0.5)
# fc7
mul7 = ad.matmul_op(drop6, filters7)
add7 = ad.add_op(mul7, biases7)
bn7 = ad.fullybn_forward_op(add7, "NCHW")
fc7 = ad.fullyactivation_forward_op(bn7, "NCHW", "relu")
drop7 = ad.fullydropout_forward_op(fc7, "NCHW", 0.5)
#fc8
mul8 = ad.matmul_op(drop7, filters8)
add8 = ad.add_op(mul8, biases8)
fc8 = ad.fullyactivation_forward_op(add8, "NCHW", "softmax")
loss = ad.l2loss_op(fc8, y_)
grad = ad.gradients(loss, [filters1_1, filters1_2, filters2_1, filters2_2, filters3_1, filters3_2, filters3_3
, filters4_1, filters4_2, filters4_3, filters5_1, filters5_2, filters5_3
, filters6, filters7])
executor = ad.Executor([grad[0], grad[1], grad[2], grad[3], grad[4], grad[5], grad[6], grad[7], grad[8], grad[9]
, grad[10], grad[11], grad[12], grad[13], grad[14], loss, y_], ctx=ctx)
aph = 1.0e-6
for i in range(20):
select = random.randint(0, n-1)
tmp_x_val = x_val[select]
tmp_x_val = np.expand_dims(tmp_x_val, 0)
tmp_y_val = y_val[select]
tmp_y_val = np.expand_dims(tmp_y_val, 0)
grad_val = executor.run(
feed_dict={inputs: tmp_x_val, y_: tmp_y_val
, filters1_1: filters_val[0], filters1_2: filters_val[1], filters2_1: filters_val[2], filters2_2: filters_val[3]
, filters3_1: filters_val[4], filters3_2: filters_val[5], filters3_3: filters_val[6]
, filters4_1: filters_val[7], filters4_2: filters_val[8], filters4_3: filters_val[9]
, filters5_1: filters_val[10], filters5_2: filters_val[11], filters5_3: filters_val[12]
, filters6: filters_val[13], filters7: filters_val[14], filters8: filters_val[15]
, biases6: biases_val[0], biases7: biases_val[1], biases8: biases_val[2]})
for i in range(14):
sgd_update_gpu(filters_val[i], grad_val[i], aph)
print(filters_val[0].asnumpy())
return filters_val
def ResNet152(inputs, n_class):
X = ad.Placeholder("X")
y_ = ad.Placeholder("y_")
W1 = ad.Variable("W1")
W6 = ad.Variable("W6")
b6 = ad.Variable("b6")
W7 = ad.Variable("W7")
b7 = ad.Variable("b7")
keep_prob = ad.Placeholder("keep_prob")
#conv1
conv1 = ad.convolution_2d_forward_op(X, W1, "NCHW", "VALID", 2, 2)
bn1 = ad.bn_forward_op(conv1, "NCHW", "pre_activation")
act1 = ad.activation_forward_op(bn1, "NCHW", "relu")
pool1 = ad.pooling_2d_forward_op(act1, "NCHW", "max", 0, 0, 2, 2, 3, 3)
#conv2_x
conv2, dict2 = convolutional_block(inputs=pool1,
kernel_size=3,
in_filter=64,
out_filters=[64, 64, 256],
block_name="2a",
stride=1)
iden2_1, dict2_1 = identity_block(inputs=conv2,
kernel_size=3,
in_filter=256,
out_filters=[64, 64, 256],
block_name="2b",
stride=1)
iden2_2, dict2_2 = identity_block(iden2_1, 3, 256, [64, 64, 256], "2c", 1)
#conv3_x
conv3, dict3 = convolutional_block(iden2_2, 3, 256, [128, 128, 512], "3a",
1)
iden3_1, dict3_1 = identity_block(conv3, 3, 512, [128, 128, 512], "3b", 1)
iden3_2, dict3_2 = identity_block(iden3_1, 3, 512, [128, 128, 512], "3c",
1)
iden3_3, dict3_3 = identity_block(iden3_2, 3, 512, [128, 128, 512], "3d",
1)
iden3_4, dict3_4 = identity_block(iden3_3, 3, 512, [128, 128, 512], "3e",
1)
iden3_5, dict3_5 = identity_block(iden3_4, 3, 512, [128, 128, 512], "3f",
1)
iden3_6, dict3_6 = identity_block(iden3_5, 3, 512, [128, 128, 512], "3g",
1)
iden3_7, dict3_7 = identity_block(iden3_6, 3, 512, [128, 128, 512], "3h",
1)
#conv4_x
conv4, dict4 = convolutional_block(iden3_7, 3, 512, [256, 256, 1024], "4a",
1)
iden4_1, dict4_1 = identity_block(conv4, 3, 1024, [256, 256, 1024], "4b",
1)
iden4_2, dict4_2 = identity_block(iden4_1, 3, 1024, [256, 256, 1024], "4c",
1)
iden4_3, dict4_3 = identity_block(iden4_2, 3, 1024, [256, 256, 1024], "4d",
1)
iden4_4, dict4_4 = identity_block(iden4_3, 3, 1024, [256, 256, 1024], "4e",
1)
iden4_5, dict4_5 = identity_block(iden4_4, 3, 1024, [256, 256, 1024], "4f",
1)
iden4_6, dict4_6 = identity_block(iden4_5, 3, 1024, [256, 256, 1024], "4f",
1)
iden4_7, dict4_7 = identity_block(iden4_6, 3, 1024, [256, 256, 1024], "4f",
1)
iden4_8, dict4_8 = identity_block(iden4_7, 3, 1024, [256, 256, 1024], "4f",
1)
iden4_9, dict4_9 = identity_block(iden4_8, 3, 1024, [256, 256, 1024], "4f",
1)
iden4_10, dict4_10 = identity_block(iden4_9, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_11, dict4_11 = identity_block(iden4_10, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_12, dict4_12 = identity_block(iden4_11, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_13, dict4_13 = identity_block(iden4_12, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_14, dict4_14 = identity_block(iden4_13, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_15, dict4_15 = identity_block(iden4_14, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_16, dict4_16 = identity_block(iden4_15, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_17, dict4_17 = identity_block(iden4_16, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_18, dict4_18 = identity_block(iden4_17, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_19, dict4_19 = identity_block(iden4_18, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_20, dict4_20 = identity_block(iden4_19, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_21, dict4_21 = identity_block(iden4_20, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_22, dict4_22 = identity_block(iden4_21, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_23, dict4_23 = identity_block(iden4_22, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_24, dict4_24 = identity_block(iden4_23, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_25, dict4_25 = identity_block(iden4_24, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_26, dict4_26 = identity_block(iden4_25, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_27, dict4_27 = identity_block(iden4_26, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_28, dict4_28 = identity_block(iden4_27, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_29, dict4_29 = identity_block(iden4_28, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_30, dict4_30 = identity_block(iden4_29, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_31, dict4_31 = identity_block(iden4_30, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_32, dict4_32 = identity_block(iden4_31, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_33, dict4_33 = identity_block(iden4_32, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_34, dict4_34 = identity_block(iden4_33, 3, 1024, [256, 256, 1024],
"4f", 1)
iden4_35, dict4_35 = identity_block(iden4_34, 3, 1024, [256, 256, 1024],
"4f", 1)
#conv5_x
conv5, dict5 = convolutional_block(iden4_35, 3, 1024, [512, 512, 2048],
"5a", 1)
iden5_1, dict5_1 = identity_block(conv5, 3, 2048, [512, 512, 2048], "5b",
1)
iden5_2, dict5_2 = identity_block(iden5_1, 3, 2048, [512, 512, 2048], "5c",
1)
pool5 = ad.pooling_2d_forward_op(iden5_2, "NCHW", "mean", 0, 0, 1, 1, 2, 2)
pool5_flat = ad.flatten_op(pool5)
mul6 = ad.matmul_op(pool5_flat, W6)
add6 = ad.add_op(mul6, b6)
act6 = ad.fullyactivation_forward_op(add6, "NCHW", "relu")
drop_out = ad.fullydropout_forward_op(act6, "NCHW", keep_prob)
mul7 = ad.matmul_op(drop_out, W7)
add7 = ad.add_op(mul7, b7)
act7 = ad.fullyactivation_forward_op(add7, "NCHW", "softmax")
loss = ad.softmaxcrossentropy_op(act7, y_)
X_val = np.random.normal(0, 0.5, (10, 3, 230, 230))
W1_val = np.random.normal(0, 0.5, (64, 3, 7, 7))
W6_val = np.random.normal(0, 0.5, (7 * 7 * 2048, 50))
b6_val = np.random.normal(0, 0.5, (10, 50))
W7_val = np.random.normal(0, 0.5, (50, 6))
b7_val = np.random.normal(0, 0.5, (10, 6))
y_val = np.random.normal(0, 0.5, (10, 6))
aph = 0.001
t = train.Adam_minimize(loss, aph)
feed_dict = {W1: W1_val, W6: W6_val, W7: W7_val, b6: b6_val, b7: b7_val}
feed_dict.update(dict2)
feed_dict.update(dict2_1)
feed_dict.update(dict2_2)
feed_dict.update(dict3)
feed_dict.update(dict3_1)
feed_dict.update(dict3_2)
feed_dict.update(dict3_3)
feed_dict.update(dict3_4)
feed_dict.update(dict3_5)
feed_dict.update(dict3_6)
feed_dict.update(dict3_7)
feed_dict.update(dict4)
feed_dict.update(dict4_1)
feed_dict.update(dict4_2)
feed_dict.update(dict4_3)
feed_dict.update(dict4_4)
feed_dict.update(dict4_5)
feed_dict.update(dict4_6)
feed_dict.update(dict4_7)
feed_dict.update(dict4_8)
feed_dict.update(dict4_9)
feed_dict.update(dict4_10)
feed_dict.update(dict4_11)
feed_dict.update(dict4_12)
feed_dict.update(dict4_13)
feed_dict.update(dict4_14)
feed_dict.update(dict4_15)
feed_dict.update(dict4_16)
feed_dict.update(dict4_17)
feed_dict.update(dict4_18)
feed_dict.update(dict4_19)
feed_dict.update(dict4_20)
feed_dict.update(dict4_21)
feed_dict.update(dict4_22)
feed_dict.update(dict4_23)
feed_dict.update(dict4_24)
feed_dict.update(dict4_25)
feed_dict.update(dict4_26)
feed_dict.update(dict4_27)
feed_dict.update(dict4_28)
feed_dict.update(dict4_29)
feed_dict.update(dict4_30)
feed_dict.update(dict4_31)
feed_dict.update(dict4_32)
feed_dict.update(dict4_33)
feed_dict.update(dict4_34)
feed_dict.update(dict4_35)
feed_dict.update(dict5)
feed_dict.update(dict5_1)
feed_dict.update(dict5_2)
# t.init_Variable(feed_dict)
# t.run({X: X_val, y_: y_val})
# print(t.get_Variable_node_to_val_map()[W1_val].asnumpy())
list = []
for key in feed_dict.keys():
list.append(key)
executor = ad.Executor(list, ctx=ndarray.gpu(0))
executor.run(feed_dict)
def ResNet50(inputs, n_class):
X = ad.Placeholder("X")
y_ = ad.Placeholder("y_")
W1 = ad.Variable("W1")
W6 = ad.Variable("W6")
b6 = ad.Variable("b6")
W7 = ad.Variable("W7")
b7 = ad.Variable("b7")
keep_prob = ad.Placeholder("keep_prob")
#conv1
conv1 = ad.convolution_2d_forward_op(X, W1, "NCHW", "VALID", 2, 2)
bn1 = ad.bn_forward_op(conv1, "NCHW", "pre_activation")
act1 = ad.activation_forward_op(bn1, "NCHW", "relu")
pool1 = ad.pooling_2d_forward_op(act1, "NCHW", "max", 0, 0, 2, 2, 3, 3)
#conv2_x
conv2, dict2 = convolutional_block(inputs=pool1,
kernel_size=3,
in_filter=64,
out_filters=[64, 64, 256],
block_name="2a",
stride=1)
iden2_1, dict2_1 = identity_block(inputs=conv2,
kernel_size=3,
in_filter=256,
out_filters=[64, 64, 256],
block_name="2b",
stride=1)
iden2_2, dict2_2 = identity_block(iden2_1, 3, 256, [64, 64, 256], "2c", 1)
#conv3_x
conv3, dict3 = convolutional_block(iden2_2, 3, 256, [128, 128, 512], "3a",
1)
iden3_1, dict3_1 = identity_block(conv3, 3, 512, [128, 128, 512], "3b", 1)
iden3_2, dict3_2 = identity_block(iden3_1, 3, 512, [128, 128, 512], "3c",
1)
iden3_3, dict3_3 = identity_block(iden3_2, 3, 512, [128, 128, 512], "3d",
1)
#conv4_x
conv4, dict4 = convolutional_block(iden3_3, 3, 512, [256, 256, 1024], "4a",
1)
iden4_1, dict4_1 = identity_block(conv4, 3, 1024, [256, 256, 1024], "4b",
1)
iden4_2, dict4_2 = identity_block(iden4_1, 3, 1024, [256, 256, 1024], "4c",
1)
iden4_3, dict4_3 = identity_block(iden4_2, 3, 1024, [256, 256, 1024], "4d",
1)
iden4_4, dict4_4 = identity_block(iden4_3, 3, 1024, [256, 256, 1024], "4e",
1)
iden4_5, dict4_5 = identity_block(iden4_4, 3, 1024, [256, 256, 1024], "4f",
1)
#conv5_x
conv5, dict5 = convolutional_block(iden4_5, 3, 1024, [512, 512, 2048],
"5a", 1)
iden5_1, dict5_1 = identity_block(conv5, 3, 2048, [512, 512, 2048], "5b",
1)
iden5_2, dict5_2 = identity_block(iden5_1, 3, 2048, [512, 512, 2048], "5c",
1)
pool5 = ad.pooling_2d_forward_op(iden5_2, "NCHW", "mean", 0, 0, 1, 1, 2, 2)
pool5_flat = ad.flatten_op(pool5)
mul6 = ad.matmul_op(pool5_flat, W6)
add6 = ad.add_op(mul6, b6)
act6 = ad.fullyactivation_forward_op(add6, "NCHW", "relu")
drop_out = ad.fullydropout_forward_op(act6, "NCHW", keep_prob)
mul7 = ad.matmul_op(drop_out, W7)
add7 = ad.add_op(mul7, b7)
act7 = ad.fullyactivation_forward_op(add7, "NCHW", "softmax")
loss = ad.softmaxcrossentropy_op(act7, y_)
X_val = np.random.normal(0, 0.5, (10, 3, 230, 230))
W1_val = np.random.normal(0, 0.5, (64, 3, 7, 7))
W6_val = np.random.normal(0, 0.5, (7 * 7 * 2048, 50))
b6_val = np.random.normal(0, 0.5, (10, 50))
W7_val = np.random.normal(0, 0.5, (50, 6))
b7_val = np.random.normal(0, 0.5, (10, 6))
y_val = np.random.normal(0, 0.5, (10, 6))
feed_dict = {W1: W1_val, W6: W6_val, W7: W7_val, b6: b6_val, b7: b7_val}
feed_dict.update(dict2)
feed_dict.update(dict2_1)
feed_dict.update(dict2_2)
feed_dict.update(dict3)
feed_dict.update(dict3_1)
feed_dict.update(dict3_2)
feed_dict.update(dict3_3)
feed_dict.update(dict4)
feed_dict.update(dict4_1)
feed_dict.update(dict4_2)
feed_dict.update(dict4_3)
feed_dict.update(dict4_4)
feed_dict.update(dict4_5)
feed_dict.update(dict5)
feed_dict.update(dict5_1)
feed_dict.update(dict5_2)
time.sleep(5)
list = []
for key in feed_dict.keys():
list.append(key)
executor = ad.Executor(list, ctx=ndarray.gpu(0))
executor.run(feed_dict)
time.sleep(5)
def vgg16(num_epochs=10):
n_class = 10
X = ad.Placeholder("inputs")
y_ = ad.Placeholder("y_")
filters1_1 = ad.Variable("filters1_1")
filters1_2 = ad.Variable("filters1_2")
filters2_1 = ad.Variable("filters2_1")
filters2_2 = ad.Variable("filters2_2")
filters3_1 = ad.Variable("filters3_1")
filters3_2 = ad.Variable("filters3_2")
filters3_3 = ad.Variable("filters3_3")
filters4_1 = ad.Variable("filters4_1")
filters4_2 = ad.Variable("filters4_2")
filters4_3 = ad.Variable("filters4_3")
filters5_1 = ad.Variable("filters5_1")
filters5_2 = ad.Variable("filters5_2")
filters5_3 = ad.Variable("filters5_3")
filters6 = ad.Variable("filters6")
filters7 = ad.Variable("filters7")
filters8 = ad.Variable("filters8")
b1_1 = ad.Variable("b1_1")
b1_2 = ad.Variable("b1_2")
b2_1 = ad.Variable("b2_1")
b2_2 = ad.Variable("b2_2")
b3_1 = ad.Variable("b3_1")
b3_2 = ad.Variable("b3_2")
b3_3 = ad.Variable("b3_3")
b4_1 = ad.Variable("b4_1")
b4_2 = ad.Variable("b4_2")
b4_3 = ad.Variable("b4_3")
b5_1 = ad.Variable("b5_1")
b5_2 = ad.Variable("b5_2")
b5_3 = ad.Variable("b5_3")
b6 = ad.Variable("b6")
b7 = ad.Variable("b7")
b8 = ad.Variable("b8")
# conv 1
conv1_1 = ad.conv2withbias(X, filters1_1, b1_1, "NCHW", "SAME", 1, 1)
bn1_1 = ad.bn_forward_op(conv1_1, "NCHW", "pre_activation")
act1_1 = ad.activation_forward_op(bn1_1, "NCHW", "relu")
conv1_2 = ad.conv2withbias(act1_1, filters1_2, b1_2, "NCHW", "SAME", 1, 1)
bn1_2 = ad.bn_forward_op(conv1_2, "NCHW", "pre_activation")
act1_2 = ad.activation_forward_op(bn1_2, "NCHW", "relu")
pool1 = ad.pooling_2d_forward_op(act1_2, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 2
conv2_1 = ad.conv2withbias(pool1, filters2_1, b2_1, "NCHW", "SAME", 1, 1)
bn2_1 = ad.bn_forward_op(conv2_1, "NCHW", "pre_activation")
act2_1 = ad.activation_forward_op(bn2_1, "NCHW", "relu")
conv2_2 = ad.conv2withbias(act2_1, filters2_2, b2_2, "NCHW", "SAME", 1, 1)
bn2_2 = ad.bn_forward_op(conv2_2, "NCHW", "pre_activation")
act2_2 = ad.activation_forward_op(bn2_2, "NCHW", "relu")
pool2 = ad.pooling_2d_forward_op(act2_2, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 3
conv3_1 = ad.conv2withbias(pool2, filters3_1, b3_1, "NCHW", "SAME", 1, 1)
bn3_1 = ad.bn_forward_op(conv3_1, "NCHW", "pre_activation")
act3_1 = ad.activation_forward_op(bn3_1, "NCHW", "relu")
conv3_2 = ad.conv2withbias(act3_1, filters3_2, b3_2, "NCHW", "SAME", 1, 1)
bn3_2 = ad.bn_forward_op(conv3_2, "NCHW", "pre_activation")
act3_2 = ad.activation_forward_op(bn3_2, "NCHW", "relu")
conv3_3 = ad.conv2withbias(act3_2, filters3_3, b3_3, "NCHW", "SAME", 1, 1)
bn3_3 = ad.bn_forward_op(conv3_3, "NCHW", "pre_activation")
act3_3 = ad.activation_forward_op(bn3_3, "NCHW", "relu")
pool3 = ad.pooling_2d_forward_op(act3_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 4
conv4_1 = ad.conv2withbias(pool3, filters4_1, b4_1, "NCHW", "SAME", 1, 1)
bn4_1 = ad.bn_forward_op(conv4_1, "NCHW", "pre_activation")
act4_1 = ad.activation_forward_op(bn4_1, "NCHW", "relu")
conv4_2 = ad.conv2withbias(act4_1, filters4_2, b4_2, "NCHW", "SAME", 1, 1)
bn4_2 = ad.bn_forward_op(conv4_2, "NCHW", "pre_activation")
act4_2 = ad.activation_forward_op(bn4_2, "NCHW", "relu")
conv4_3 = ad.conv2withbias(act4_2, filters4_3, b4_3, "NCHW", "SAME", 1, 1)
bn4_3 = ad.bn_forward_op(conv4_3, "NCHW", "pre_activation")
act4_3 = ad.activation_forward_op(bn4_3, "NCHW", "relu")
pool4 = ad.pooling_2d_forward_op(act4_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# conv 5
conv5_1 = ad.conv2withbias(pool4, filters5_1, b5_1, "NCHW", "SAME", 1, 1)
bn5_1 = ad.bn_forward_op(conv5_1, "NCHW", "pre_activation")
act5_1 = ad.activation_forward_op(bn5_1, "NCHW", "relu")
conv5_2 = ad.conv2withbias(act5_1, filters5_2, b5_2, "NCHW", "SAME", 1, 1)
bn5_2 = ad.bn_forward_op(conv5_2, "NCHW", "pre_activation")
act5_2 = ad.activation_forward_op(bn5_2, "NCHW", "relu")
conv5_3 = ad.conv2withbias(act5_2, filters5_3, b5_3, "NCHW", "SAME", 1, 1)
bn5_3 = ad.bn_forward_op(conv5_3, "NCHW", "pre_activation")
act5_3 = ad.activation_forward_op(bn5_3, "NCHW", "relu")
pool5 = ad.pooling_2d_forward_op(act5_3, "NCHW", "max", 0, 0, 2, 2, 2, 2)
# fc6
pool5_flat = ad.flatten_op(pool5)
fc6 = ad.dense(pool5_flat, filters6, b6)
bn6 = ad.fullybn_forward_op(fc6, "NCHW")
act6 = ad.fullyactivation_forward_op(bn6, "NCHW", "relu")
drop6 = ad.fullydropout_forward_op(act6, "NCHW", 0.5)
# fc7
fc7 = ad.dense(drop6, filters7, b7)
bn7 = ad.fullybn_forward_op(fc7, "NCHW")
act7 = ad.fullyactivation_forward_op(bn7, "NCHW", "relu")
drop7 = ad.fullydropout_forward_op(act7, "NCHW", 0.5)
#fc8
fc8 = ad.dense(drop7, filters8, b8)
bn8 = ad.fullybn_forward_op(fc8, "NCHW")
act8 = ad.fullyactivation_forward_op(bn8, "NCHW", "softmax")
loss = ad.softmaxcrossentropy_op(bn8, y_)
# grad = ad.gradients(loss, [filters1_1, filters1_2, filters2_1, filters2_2, filters3_1, filters3_2, filters3_3
# , filters4_1, filters4_2, filters4_3, filters5_1, filters5_2, filters5_3
# , filters6, filters7])
# executor = ad.Executor([grad[0], grad[1], grad[2], grad[3], grad[4], grad[5], grad[6], grad[7], grad[8], grad[9]
# , grad[10], grad[11], grad[12], grad[13], grad[14], loss, y_], ctx=ctx)
train_x, train_y, test_x, test_y = prepare_data()
n_train_batches = train_x.shape[0] // batch_size
n_test_batches = test_x.shape[0] // batch_size
X_val = np.empty(shape=(batch_size, 3, 32, 32), dtype=np.float32)
y_val = np.empty(shape=(batch_size, n_class), dtype=np.float32)
filters_val = [np.random.normal(0.0, 0.1, (64, 3, 3, 3))]
filters_val.append(np.random.normal(0.0, 0.1, (64, 64, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (128, 64, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (128, 128, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (256, 128, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (256, 256, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (256, 256, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 256, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 512, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 512, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 512, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 512, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512, 512, 3, 3)))
filters_val.append(np.random.normal(0.0, 0.1, (512 * 1 * 1, 4096)))
filters_val.append(np.random.normal(0.0, 0.1, (4096, 4096)) * 0.001)
filters_val.append(np.random.normal(0.0, 0.1, (4096, n_class)) * 0.001)
b_val = [np.ones(64) * 0.1]
b_val.append(np.ones(64) * 0.1)
b_val.append(np.ones(128) * 0.1)
b_val.append(np.ones(128) * 0.1)
b_val.append(np.ones(256) * 0.1)
b_val.append(np.ones(256) * 0.1)
b_val.append(np.ones(256) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(512) * 0.1)
b_val.append(np.ones(4096) * 0.1)
b_val.append(np.ones(4096) * 0.1)
b_val.append(np.ones(n_class) * 0.1)
# ctx = ndarray.gpu(0)
# for i in range(16):
# filters_val[i] = ndarray.array(filters_val[i], ctx)
# for i in range(16):
# b_val[i] = ndarray.array(b_val[i], ctx)
aph = 0.0001
t = train.Adam_minimize(loss, aph)
t.init_Variable({
filters1_1: filters_val[0],
filters1_2: filters_val[1],
filters2_1: filters_val[2],
filters2_2: filters_val[3],
filters3_1: filters_val[4],
filters3_2: filters_val[5],
filters3_3: filters_val[6],
filters4_1: filters_val[7],
filters4_2: filters_val[8],
filters4_3: filters_val[9],
filters5_1: filters_val[10],
filters5_2: filters_val[11],
filters5_3: filters_val[12],
filters6: filters_val[13],
filters7: filters_val[14],
filters8: filters_val[15],
b1_1: b_val[0],
b1_2: b_val[1],
b2_1: b_val[2],
b2_2: b_val[3],
b3_1: b_val[4],
b3_2: b_val[5],
b3_3: b_val[6],
b4_1: b_val[7],
b4_2: b_val[8],
b4_3: b_val[9],
b5_1: b_val[10],
b5_2: b_val[11],
b5_3: b_val[12],
b6: b_val[13],
b7: b_val[14],
b8: b_val[15]
})
for i in range(num_epochs):
print("epoch %d" % i)
X_val = np.empty(shape=(0, 3, 32, 32), dtype=np.float32)
y_val = np.empty(shape=(0, n_class), dtype=np.float32)
select = random.randint(0, n_train_batches)
# X_val = train_x[select]
# y_val = train_y[select]
for j in range(batch_size):
select = random.randint(0, n_train_batches)
temp_train_x = train_x[select]
temp_train_x = temp_train_x[np.newaxis, :]
temp_train_y = train_y[select]
temp_train_y = temp_train_y[np.newaxis, :]
X_val = np.append(X_val, temp_train_x, axis=0)
y_val = np.append(y_val, temp_train_y, axis=0)
# print("train", X_val.shape)
# print("train", y_val.shape)
t.run({X: X_val, y_: y_val})
if (i + 1) % 30 == 0:
X_test_val = np.empty(shape=(100, 3, 32, 32), dtype=np.float32)
y_test_val = np.empty(shape=(100, n_class), dtype=np.float32)
correct_predictions = []
for minibatch_index in range(n_test_batches):
minibatch_start = minibatch_index * batch_size
minibatch_end = (minibatch_index + 1) * batch_size
X_test_val[:] = test_x[minibatch_start:minibatch_end]
y_test_val[:] = test_y[minibatch_start:minibatch_end]
# print(X_test_val.shape)
# print(y_test_val.shape)
feed_dict = {X: X_test_val, y_: y_test_val}
dict_Variable = t.get_Variable_node_to_val_map()
feed_dict.update(dict_Variable)
y_predicted = t.run_get_nodelist_once(feed_dict,
[act8])[act8].asnumpy()
correct_prediction = np.equal(np.argmax(y_test_val, 1),
np.argmax(y_predicted,
1)).astype(np.float)
correct_predictions.extend(correct_prediction)
accuracy = np.mean(correct_predictions)
print("validation set accuracy=%f" % accuracy)
return filters_val
